9916 lines
291 KiB
JSON
9916 lines
291 KiB
JSON
{
|
|
"blocks": [
|
|
{
|
|
"action_pairs": 12,
|
|
"coverage_at_end": {
|
|
"admitted_fraction_of_total": {
|
|
"distinct_n": 21,
|
|
"max": 0.12295081967213115,
|
|
"median": 0.08985507246376812,
|
|
"min": 0.04918032786885246,
|
|
"n": 36
|
|
},
|
|
"completed_fraction_of_total": {
|
|
"distinct_n": 30,
|
|
"max": 0.09361702127659574,
|
|
"median": 0.050437283852088316,
|
|
"min": 0.005586592178770949,
|
|
"n": 36
|
|
}
|
|
},
|
|
"efficacy": {
|
|
"feasibility_transitions": {
|
|
"false->false": 3,
|
|
"false->true": 6,
|
|
"true->true": 3
|
|
},
|
|
"minimum_balanced_accuracy": 0.75,
|
|
"minimum_delta_over_best_outcome": 0.15,
|
|
"outcome_delta": {
|
|
"best_accuracy": 0.75,
|
|
"best_balanced_accuracy": 0.75,
|
|
"best_feature": "ttft_max_over_slo_max",
|
|
"features": {
|
|
"admitted_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"admitted_input_tokens_mean_over_limit": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_fail_fraction_of_total": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_over_admitted": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.4166666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.051724137931034475,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.09339080459770116,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.018181818181818188,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"completed_pass_rate": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"outstanding_over_admitted": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.4166666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.051724137931034475,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.09339080459770113,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.01818181818181816,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"tpot_max_over_slo": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.000601708027350864,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.000601708027350864,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.015623848188639142,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_mean_over_slo": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0010320465723873684,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.009580791853053022,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0002360795908454838,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"ttft_max_over_slo_max": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.0072977563346891365,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.021249613758603417,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.007412010668000825,
|
|
"train_balanced_accuracy": 0.875
|
|
}
|
|
]
|
|
},
|
|
"ttft_mean_over_slo_max": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.0008523871757378475,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.001080823830574982,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.006279735096096474,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_delta": {
|
|
"best_accuracy": 0.6666666666666666,
|
|
"best_balanced_accuracy": 0.6666666666666666,
|
|
"best_feature": "queue_waiting_mean",
|
|
"features": {
|
|
"decode_batch_size.mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.5847120251326046,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.1503494070398912,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.759913509761732,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"graph_padding_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0003962349313431858,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0006270962078854048,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.003423657452078254,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"kv_usage_mean": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00016329402239600305,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00013516820682281123,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.00013516820682281123,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"prefill_token_fraction": {
|
|
"accuracy": 0.3333333333333333,
|
|
"balanced_accuracy": 0.3333333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0003659931906908054,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0003153433269442174,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.002846241748737599,
|
|
"train_balanced_accuracy": 1.0
|
|
}
|
|
]
|
|
},
|
|
"queue_running_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 1.2024471918333335,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.18325025108333315,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 1.2024471918333335,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_waiting_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -1.6539956596666667,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.009266072583333333,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -1.6539956596666667,
|
|
"train_balanced_accuracy": 0.875
|
|
}
|
|
]
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 1.5833333333333357,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 4.416666666666664,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -19.583333333333332,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_qualifying_features": []
|
|
},
|
|
"end_fraction": 0.1,
|
|
"end_s": 6.0,
|
|
"qualifying_response_features": [],
|
|
"repeat_pairs": 24,
|
|
"response_statistics": {
|
|
"batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.587816632458785,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 10.944163150492265,
|
|
"min": -0.9368137011239002,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.1464621283214008,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 4.013437734353163,
|
|
"repeat_abs_p95": 8.899704585700528,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 9.092698412698415,
|
|
"min": -9.933716475095787,
|
|
"n": 24
|
|
}
|
|
},
|
|
"batch_tokens.mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 7.901570082144033,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 564.520698546648,
|
|
"min": -8.58386322115848,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.05293318721992956,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 149.27440604160446,
|
|
"repeat_abs_p95": 426.8141407023969,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 449.2116931216931,
|
|
"min": -584.6987867177522,
|
|
"n": 24
|
|
}
|
|
},
|
|
"decode_batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.582158519407685,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 10.735314111579934,
|
|
"min": -0.9288831363334129,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.14646507698937813,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 3.9747257938484823,
|
|
"repeat_abs_p95": 8.770067452024344,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 8.954735449735448,
|
|
"min": -9.735983397190292,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_full_share": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.005042559238173261,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.008374384236453203,
|
|
"min": -0.10829817158931077,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.12202178454078347,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.04132507369196753,
|
|
"repeat_abs_p95": 0.08139543508103343,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.12107279693486583,
|
|
"min": -0.0822222222222222,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_none_share": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.008375994890060249,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.09274496015002344,
|
|
"min": -0.01986863711001642,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.75,
|
|
"negative": 9,
|
|
"positive": 3,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.22398385207837773,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.037395530134597704,
|
|
"repeat_abs_p95": 0.07586812708364432,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.07631077509373656,
|
|
"min": -0.10941890166028097,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_padding_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0011703901376772582,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.006703036588479568,
|
|
"min": -0.0015820428864207082,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.75,
|
|
"negative": 3,
|
|
"positive": 9,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.29946486376659237,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.003908271985422232,
|
|
"repeat_abs_p95": 0.008187883039017584,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.007141591578301508,
|
|
"min": -0.009201187757062074,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_end_minus_start": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.006820888350695331,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.140893666867539,
|
|
"min": -0.005054735620165451,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.8333333333333334,
|
|
"negative": 2,
|
|
"positive": 10,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6661896437498689,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.010238658638254572,
|
|
"repeat_abs_p95": 0.1277225497052652,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.13402937356379263,
|
|
"min": -0.1168947946847837,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_max": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.007604372340016818,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.06684057734774362,
|
|
"min": 0.00016629145416469093,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.3770466591391387,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.020168252802926,
|
|
"repeat_abs_p95": 0.11084274153262068,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.12803476870816266,
|
|
"min": -0.11234888600259774,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_mean": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0014831270777658032,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.02895016189128262,
|
|
"min": -0.00029784868681948656,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 4,
|
|
"positive": 8,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.12120157645159224,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.01223686292857884,
|
|
"repeat_abs_p95": 0.07958153755957757,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.08820147130614941,
|
|
"min": -0.08208138515366167,
|
|
"n": 24
|
|
}
|
|
},
|
|
"preemptions": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0,
|
|
"action_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.0,
|
|
"negative": 0,
|
|
"positive": 0,
|
|
"zero": 12
|
|
},
|
|
"effect_to_repeat_median": 0.0,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0,
|
|
"repeat_abs_p95": 0.0,
|
|
"repeat_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 24
|
|
}
|
|
},
|
|
"prefill_token_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0028294297587345696,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.013691652669458043,
|
|
"min": -0.002949832419021914,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.44362661948582977,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.006377953067861264,
|
|
"repeat_abs_p95": 0.031272369907012425,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.036461564751038456,
|
|
"min": -0.018166035027340932,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_running_mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.50741128125,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 15.614046583666667,
|
|
"min": -0.30207624649999953,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.12962585804860988,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 3.9144294887499997,
|
|
"repeat_abs_p95": 7.101594725466666,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 6.513803891666667,
|
|
"min": -13.5536487565,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_waiting_mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.24895167858333334,
|
|
"action_delta": {
|
|
"distinct_n": 10,
|
|
"max": 0.0,
|
|
"min": -8.040727011166666,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 9,
|
|
"positive": 0,
|
|
"zero": 3
|
|
},
|
|
"effect_to_repeat_median": 1.1132189691456902,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.22363226416666668,
|
|
"repeat_abs_p95": 2.296808863758333,
|
|
"repeat_delta": {
|
|
"distinct_n": 16,
|
|
"max": 2.3311656403333334,
|
|
"min": -5.974843045,
|
|
"n": 24
|
|
}
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 3.249999999999993,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 10.5,
|
|
"min": -30.999999999999993,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.1065573770491801,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 30.5,
|
|
"repeat_abs_p95": 59.92499999999999,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 59.5,
|
|
"min": -64.5,
|
|
"n": 24
|
|
}
|
|
}
|
|
},
|
|
"sanity": {
|
|
"action_pairs": 12,
|
|
"invariants": {
|
|
"efficacy_label_balance": true,
|
|
"expected_action_pair_count": true,
|
|
"expected_repeat_pair_count": true,
|
|
"finite_deltas": true,
|
|
"matched_action_request_hashes": true,
|
|
"probabilities_bounded": true
|
|
},
|
|
"red_flags": [],
|
|
"repeat_pairs": 24,
|
|
"trials": 36
|
|
},
|
|
"start_s": 0.0
|
|
},
|
|
{
|
|
"action_pairs": 12,
|
|
"coverage_at_end": {
|
|
"admitted_fraction_of_total": {
|
|
"distinct_n": 22,
|
|
"max": 0.20535714285714285,
|
|
"median": 0.18235294117647058,
|
|
"min": 0.1564245810055866,
|
|
"n": 36
|
|
},
|
|
"completed_fraction_of_total": {
|
|
"distinct_n": 31,
|
|
"max": 0.180327868852459,
|
|
"median": 0.14411764705882352,
|
|
"min": 0.08933333333333335,
|
|
"n": 36
|
|
}
|
|
},
|
|
"efficacy": {
|
|
"feasibility_transitions": {
|
|
"false->false": 3,
|
|
"false->true": 6,
|
|
"true->true": 3
|
|
},
|
|
"minimum_balanced_accuracy": 0.75,
|
|
"minimum_delta_over_best_outcome": 0.15,
|
|
"outcome_delta": {
|
|
"best_accuracy": 0.8333333333333334,
|
|
"best_balanced_accuracy": 0.8333333333333333,
|
|
"best_feature": "ttft_max_over_slo_max",
|
|
"features": {
|
|
"admitted_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"admitted_input_tokens_mean_over_limit": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_fail_fraction_of_total": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.012666666666666666,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.008666666666666666,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.008666666666666666,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.03401360544217691,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.006172839506172811,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.04871948779511809,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_pass_rate": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.11875000000000002,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.07738095238095238,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.07738095238095238,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"outstanding_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.03401360544217688,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.006172839506172839,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.04871948779511806,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_max_over_slo": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.028440662518539606,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.02030446551608167,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.008882265987138677,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_mean_over_slo": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.0020290170113650557,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.04992359808650447,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.009324194656287296,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_max_over_slo_max": {
|
|
"accuracy": 0.8333333333333334,
|
|
"balanced_accuracy": 0.8333333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.01753702182516767,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.01753702182516767,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.018086205842943556,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_mean_over_slo_max": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.010556909058063094,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.003084423316020539,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.027367186208561683,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_delta": {
|
|
"best_accuracy": 0.75,
|
|
"best_balanced_accuracy": 0.75,
|
|
"best_feature": "queue_waiting_mean",
|
|
"features": {
|
|
"decode_batch_size.mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.22344269117369397,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.22344269117369397,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.5979430550644496,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"graph_padding_fraction": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.006014036688642874,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.003996367458148475,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.004817288488210095,
|
|
"train_balanced_accuracy": 0.875
|
|
}
|
|
]
|
|
},
|
|
"kv_usage_mean": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.0009690245761390937,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.014778075272949286,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00021729561459569828,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"prefill_token_fraction": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0004249745076460587,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0008306228170546492,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.005838187880452539,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_running_mean": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.014251387166666962,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 9.969867595499998,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.36342276333333334,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_waiting_mean": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.37299224008333337,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.09157276941666666,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.08574448658333333,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 3.75,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 3.75,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 6.833333333333329,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_qualifying_features": []
|
|
},
|
|
"end_fraction": 0.2,
|
|
"end_s": 12.0,
|
|
"qualifying_response_features": [],
|
|
"repeat_pairs": 24,
|
|
"response_statistics": {
|
|
"batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.836467208595514,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 38.01660606060606,
|
|
"min": -3.7894183532722323,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6942297063022367,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 1.2048853585521353,
|
|
"repeat_abs_p95": 9.90189234803657,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 6.213858257127487,
|
|
"min": -19.9836254925941,
|
|
"n": 24
|
|
}
|
|
},
|
|
"batch_tokens.mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 68.40350288667949,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 1287.087393939394,
|
|
"min": -109.74922437881139,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.6828036858921454,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 40.648534026959354,
|
|
"repeat_abs_p95": 252.72760702584105,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 53.09534549057537,
|
|
"min": -841.3287810843865,
|
|
"n": 24
|
|
}
|
|
},
|
|
"decode_batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.79935614604261,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 37.563515151515155,
|
|
"min": -3.7555281535670666,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6727103099461282,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 1.1882620709449565,
|
|
"repeat_abs_p95": 9.833929419939027,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 6.226902904787521,
|
|
"min": -19.76212800652263,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_full_share": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.023223408379118127,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.03622200353327565,
|
|
"min": -0.23381818181818181,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 2.5685579123999758,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.009041419026218855,
|
|
"repeat_abs_p95": 0.07343409055233649,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.08839516238619383,
|
|
"min": -0.027874161461947766,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_none_share": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.0196424445317687,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.22624242424242422,
|
|
"min": -0.043130466400287726,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 7,
|
|
"positive": 5,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.9837484419794449,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.009901681138647232,
|
|
"repeat_abs_p95": 0.07568864235997094,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.02793199167244964,
|
|
"min": -0.10324092947411331,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_padding_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0019741621395090708,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.013535143477830873,
|
|
"min": -0.0002595396932579229,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.9166666666666666,
|
|
"negative": 1,
|
|
"positive": 11,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6824854579006991,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.002892606892433317,
|
|
"repeat_abs_p95": 0.013935039231725855,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.016710257140637542,
|
|
"min": -0.0048630254047383765,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_end_minus_start": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.016292713617491184,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.06869972057736717,
|
|
"min": -0.13905774767384216,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 7,
|
|
"positive": 5,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.2163146605755397,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.013395146951350356,
|
|
"repeat_abs_p95": 0.10929110116632139,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.07238485363173153,
|
|
"min": -0.2557698071735438,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_max": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.01447565921273003,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.10932093798799158,
|
|
"min": 0.0002683613903166071,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.4049345895249048,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.010303439975540174,
|
|
"repeat_abs_p95": 0.07329653312019176,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.05015486062543706,
|
|
"min": -0.12688580277750028,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.005335907884240179,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.06835893039525985,
|
|
"min": -0.0020715594873930866,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.8333333333333334,
|
|
"negative": 2,
|
|
"positive": 10,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.4198198744789212,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.003758158327089537,
|
|
"repeat_abs_p95": 0.054511554710704666,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.05530066409194215,
|
|
"min": -0.1166293577215588,
|
|
"n": 24
|
|
}
|
|
},
|
|
"preemptions": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0,
|
|
"action_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.0,
|
|
"negative": 0,
|
|
"positive": 0,
|
|
"zero": 12
|
|
},
|
|
"effect_to_repeat_median": 0.0,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0,
|
|
"repeat_abs_p95": 0.0,
|
|
"repeat_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 24
|
|
}
|
|
},
|
|
"prefill_token_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.00639806071748561,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.01743846124119386,
|
|
"min": -0.012883247545976029,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.5860208014218505,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.010917804797990316,
|
|
"repeat_abs_p95": 0.038061500221448455,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.006259881301210357,
|
|
"min": -0.03966667403951463,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_running_mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.8894945159166667,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 39.57822979983333,
|
|
"min": -2.9770223346666658,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.4702668338949774,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 1.8914676770833334,
|
|
"repeat_abs_p95": 10.096755181974991,
|
|
"repeat_delta": {
|
|
"distinct_n": 21,
|
|
"max": 6.583703286333334,
|
|
"min": -20.7125852155,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_waiting_mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.79726731775,
|
|
"action_delta": {
|
|
"distinct_n": 11,
|
|
"max": 0.0,
|
|
"min": -30.713414816833335,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 10,
|
|
"positive": 0,
|
|
"zero": 2
|
|
},
|
|
"effect_to_repeat_median": 17.026052234021265,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.04682631691666666,
|
|
"repeat_abs_p95": 3.5462989908416667,
|
|
"repeat_delta": {
|
|
"distinct_n": 17,
|
|
"max": 3.561578917166667,
|
|
"min": -17.805020274833332,
|
|
"n": 24
|
|
}
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 13.083333333333334,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 23.66666666666667,
|
|
"min": -40.5,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 7,
|
|
"positive": 5,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.2661290322580652,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 10.333333333333329,
|
|
"repeat_abs_p95": 55.25833333333328,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 71.83333333333333,
|
|
"min": -19.333333333333343,
|
|
"n": 24
|
|
}
|
|
}
|
|
},
|
|
"sanity": {
|
|
"action_pairs": 12,
|
|
"invariants": {
|
|
"efficacy_label_balance": true,
|
|
"expected_action_pair_count": true,
|
|
"expected_repeat_pair_count": true,
|
|
"finite_deltas": true,
|
|
"matched_action_request_hashes": true,
|
|
"probabilities_bounded": true
|
|
},
|
|
"red_flags": [],
|
|
"repeat_pairs": 24,
|
|
"trials": 36
|
|
},
|
|
"start_s": 6.0
|
|
},
|
|
{
|
|
"action_pairs": 12,
|
|
"coverage_at_end": {
|
|
"admitted_fraction_of_total": {
|
|
"distinct_n": 22,
|
|
"max": 0.30654761904761907,
|
|
"median": 0.2795403407264545,
|
|
"min": 0.24313725490196078,
|
|
"n": 36
|
|
},
|
|
"completed_fraction_of_total": {
|
|
"distinct_n": 30,
|
|
"max": 0.2681564245810056,
|
|
"median": 0.22518087736972248,
|
|
"min": 0.12849162011173185,
|
|
"n": 36
|
|
}
|
|
},
|
|
"efficacy": {
|
|
"feasibility_transitions": {
|
|
"false->false": 3,
|
|
"false->true": 6,
|
|
"true->true": 3
|
|
},
|
|
"minimum_balanced_accuracy": 0.75,
|
|
"minimum_delta_over_best_outcome": 0.15,
|
|
"outcome_delta": {
|
|
"best_accuracy": 0.75,
|
|
"best_balanced_accuracy": 0.75,
|
|
"best_feature": "completed_pass_rate",
|
|
"features": {
|
|
"admitted_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"admitted_input_tokens_mean_over_limit": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_fail_fraction_of_total": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.02825325884543762,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.02266666666666667,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.040253258845437616,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.03272660415517559,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.022522522522522515,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.07358436332279389,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_pass_rate": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.14270516717325227,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.12142857142857144,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.2182462927143778,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"outstanding_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.03272660415517557,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.022522522522522515,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0735843633227939,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_max_over_slo": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.37412821213021774,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.009857117327656123,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.008882265987138677,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"tpot_mean_over_slo": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.004552157689988692,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.007167514027906635,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.006851562902858294,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_max_over_slo_max": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.017755539496041216,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.007733676747496546,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.5040019049192779,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_mean_over_slo_max": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.007216190772090309,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.007216190772090309,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.24525856553394837,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_delta": {
|
|
"best_accuracy": 0.75,
|
|
"best_balanced_accuracy": 0.75,
|
|
"best_feature": "graph_padding_fraction",
|
|
"features": {
|
|
"decode_batch_size.mean": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 1.367157621094477,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.45324800586584313,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.7719319083901095,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"graph_padding_fraction": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0006822319057190087,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0006574027141346845,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0007737011910822953,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"kv_usage_mean": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0014242600332016828,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00020601061133531957,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0014242600332016828,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"prefill_token_fraction": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.0001890219308638419,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 6.874128372463817e-05,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0012836451412676642,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_running_mean": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 2.090120409833334,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.4494322117499996,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 1.383383161750001,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_waiting_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.015561036916666665,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.015561036916666665,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.35703936641666667,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -5.666666666666664,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 3.0,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -2.6666666666666607,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_qualifying_features": []
|
|
},
|
|
"end_fraction": 0.3,
|
|
"end_s": 18.0,
|
|
"qualifying_response_features": [],
|
|
"repeat_pairs": 24,
|
|
"response_statistics": {
|
|
"batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 1.4376624963208817,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 44.66904761904762,
|
|
"min": -3.1358226741876534,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.5185498088911062,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.9467338429752983,
|
|
"repeat_abs_p95": 8.569947817628815,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 4.8797843665768195,
|
|
"min": -24.305454429328734,
|
|
"n": 24
|
|
}
|
|
},
|
|
"batch_tokens.mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 29.54235812446782,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 981.3419420915683,
|
|
"min": -40.25177564153491,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 4,
|
|
"positive": 8,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.44186157650879665,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 66.85885285135149,
|
|
"repeat_abs_p95": 331.06991793371697,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 347.23329948092976,
|
|
"min": -494.75378424985627,
|
|
"n": 24
|
|
}
|
|
},
|
|
"decode_batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 1.426379390321979,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 44.29932712215321,
|
|
"min": -3.1155318979899898,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.5217762546171796,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.9373121613602792,
|
|
"repeat_abs_p95": 8.519190735455247,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 4.8432704402515725,
|
|
"min": -24.101136871686787,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_full_share": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 0.010376892235975155,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.020290776197664284,
|
|
"min": -0.19708796988318666,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.5862533874020509,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0177003535654775,
|
|
"repeat_abs_p95": 0.05622040303840937,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.14549402823018454,
|
|
"min": -0.05811329271044907,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_none_share": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 0.015566450509702876,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.1830537562974035,
|
|
"min": -0.028434098673234293,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 7,
|
|
"positive": 5,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.9711084270914145,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.016029570000052673,
|
|
"repeat_abs_p95": 0.06874268847661695,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.07111825772963212,
|
|
"min": -0.1536373507057546,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_padding_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.0019204725324846257,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.011531473650189953,
|
|
"min": 1.5710198019351293e-05,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.109916935840515,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0017302849163485332,
|
|
"repeat_abs_p95": 0.011201032721421313,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.012042080910355725,
|
|
"min": -0.011611080181771623,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_end_minus_start": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.012368563885621575,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.1693202894846253,
|
|
"min": -0.027238473582144707,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 4,
|
|
"positive": 8,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.1201833133796462,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.01104155341174029,
|
|
"repeat_abs_p95": 0.14212743371007872,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.1535118393445899,
|
|
"min": -0.2053651713457888,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_max": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.005959053027072225,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.22224967428314468,
|
|
"min": 3.590298430544703e-05,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.49652693165856604,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.012001469904496409,
|
|
"repeat_abs_p95": 0.06026674766281165,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.06029573383954434,
|
|
"min": -0.267359376561095,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.0057476210577813765,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.18534326857338257,
|
|
"min": -0.015438462146793161,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6334301280804308,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.009073804359762887,
|
|
"repeat_abs_p95": 0.0458855138590948,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.033340684091908934,
|
|
"min": -0.24888103806759748,
|
|
"n": 24
|
|
}
|
|
},
|
|
"preemptions": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0,
|
|
"action_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.0,
|
|
"negative": 0,
|
|
"positive": 0,
|
|
"zero": 12
|
|
},
|
|
"effect_to_repeat_median": 0.0,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0,
|
|
"repeat_abs_p95": 0.0,
|
|
"repeat_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 24
|
|
}
|
|
},
|
|
"prefill_token_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.003319188436037235,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.013152637562827363,
|
|
"min": -0.02492780421592411,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.1361252231813899,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.02438334614602866,
|
|
"repeat_abs_p95": 0.07693546674256392,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.08157979537248183,
|
|
"min": -0.04610629729936333,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_running_mean": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 1.8914151395833336,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 45.162770761333334,
|
|
"min": -2.3232337786666672,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.438574132498461,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 1.3147846168333333,
|
|
"repeat_abs_p95": 6.971346508883332,
|
|
"repeat_delta": {
|
|
"distinct_n": 21,
|
|
"max": 4.9605220578333356,
|
|
"min": -20.271844710166665,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_waiting_mean": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 1.12754089475,
|
|
"action_delta": {
|
|
"distinct_n": 10,
|
|
"max": 0.0,
|
|
"min": -47.07583785533333,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 9,
|
|
"positive": 0,
|
|
"zero": 3
|
|
},
|
|
"effect_to_repeat_median": 16.245384499464624,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.06940684566666666,
|
|
"repeat_abs_p95": 5.181811176816667,
|
|
"repeat_delta": {
|
|
"distinct_n": 18,
|
|
"max": 0.6568425336666667,
|
|
"min": -7.1538662623333344,
|
|
"n": 24
|
|
}
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 7.166666666666668,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 34.0,
|
|
"min": -47.33333333333333,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.31617647058823534,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 22.666666666666668,
|
|
"repeat_abs_p95": 55.666666666666664,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 85.33333333333333,
|
|
"min": -55.66666666666667,
|
|
"n": 24
|
|
}
|
|
}
|
|
},
|
|
"sanity": {
|
|
"action_pairs": 12,
|
|
"invariants": {
|
|
"efficacy_label_balance": true,
|
|
"expected_action_pair_count": true,
|
|
"expected_repeat_pair_count": true,
|
|
"finite_deltas": true,
|
|
"matched_action_request_hashes": true,
|
|
"probabilities_bounded": true
|
|
},
|
|
"red_flags": [],
|
|
"repeat_pairs": 24,
|
|
"trials": 36
|
|
},
|
|
"start_s": 12.0
|
|
}
|
|
],
|
|
"claim_boundary": "Post-hoc corrective audit over every common replay decile. It can diagnose horizon sensitivity but cannot establish a held-out tuning claim.",
|
|
"cumulative": [
|
|
{
|
|
"action_pairs": 12,
|
|
"coverage_at_end": {
|
|
"admitted_fraction_of_total": {
|
|
"distinct_n": 21,
|
|
"max": 0.12295081967213115,
|
|
"median": 0.08985507246376812,
|
|
"min": 0.04918032786885246,
|
|
"n": 36
|
|
},
|
|
"completed_fraction_of_total": {
|
|
"distinct_n": 30,
|
|
"max": 0.09361702127659574,
|
|
"median": 0.050437283852088316,
|
|
"min": 0.005586592178770949,
|
|
"n": 36
|
|
}
|
|
},
|
|
"efficacy": {
|
|
"feasibility_transitions": {
|
|
"false->false": 3,
|
|
"false->true": 6,
|
|
"true->true": 3
|
|
},
|
|
"minimum_balanced_accuracy": 0.75,
|
|
"minimum_delta_over_best_outcome": 0.15,
|
|
"outcome_delta": {
|
|
"best_accuracy": 0.75,
|
|
"best_balanced_accuracy": 0.75,
|
|
"best_feature": "ttft_max_over_slo_max",
|
|
"features": {
|
|
"admitted_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"admitted_input_tokens_mean_over_limit": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_fail_fraction_of_total": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_over_admitted": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.4166666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.051724137931034475,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.09339080459770116,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.018181818181818188,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"completed_pass_rate": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"outstanding_over_admitted": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.4166666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.051724137931034475,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.09339080459770113,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.01818181818181816,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"tpot_max_over_slo": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.000601708027350864,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.000601708027350864,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.015623848188639142,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_mean_over_slo": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0010320465723873684,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.009580791853053022,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0002360795908454838,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"ttft_max_over_slo_max": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.0072977563346891365,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.021249613758603417,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.007412010668000825,
|
|
"train_balanced_accuracy": 0.875
|
|
}
|
|
]
|
|
},
|
|
"ttft_mean_over_slo_max": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.0008523871757378475,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.001080823830574982,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.006279735096096474,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_delta": {
|
|
"best_accuracy": 0.6666666666666666,
|
|
"best_balanced_accuracy": 0.6666666666666666,
|
|
"best_feature": "queue_waiting_mean",
|
|
"features": {
|
|
"decode_batch_size.mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.5847120251326046,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.1503494070398912,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.759913509761732,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"graph_padding_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.0003962349313431858,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0006270962078854048,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.003423657452078254,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"kv_usage_mean": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00016329402239600305,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00013516820682281123,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.00013516820682281123,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"prefill_token_fraction": {
|
|
"accuracy": 0.3333333333333333,
|
|
"balanced_accuracy": 0.3333333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0003659931906908054,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0003153433269442174,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.002846241748737599,
|
|
"train_balanced_accuracy": 1.0
|
|
}
|
|
]
|
|
},
|
|
"queue_running_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 1.2024471918333335,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.18325025108333315,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 1.2024471918333335,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_waiting_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -1.6539956596666667,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.009266072583333333,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -1.6539956596666667,
|
|
"train_balanced_accuracy": 0.875
|
|
}
|
|
]
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 1.5833333333333357,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 4.416666666666664,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -19.583333333333332,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_qualifying_features": []
|
|
},
|
|
"end_fraction": 0.1,
|
|
"end_s": 6.0,
|
|
"qualifying_response_features": [],
|
|
"repeat_pairs": 24,
|
|
"response_statistics": {
|
|
"batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.587816632458785,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 10.944163150492265,
|
|
"min": -0.9368137011239002,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.1464621283214008,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 4.013437734353163,
|
|
"repeat_abs_p95": 8.899704585700528,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 9.092698412698415,
|
|
"min": -9.933716475095787,
|
|
"n": 24
|
|
}
|
|
},
|
|
"batch_tokens.mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 7.901570082144033,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 564.520698546648,
|
|
"min": -8.58386322115848,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.05293318721992956,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 149.27440604160446,
|
|
"repeat_abs_p95": 426.8141407023969,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 449.2116931216931,
|
|
"min": -584.6987867177522,
|
|
"n": 24
|
|
}
|
|
},
|
|
"decode_batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.582158519407685,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 10.735314111579934,
|
|
"min": -0.9288831363334129,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.14646507698937813,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 3.9747257938484823,
|
|
"repeat_abs_p95": 8.770067452024344,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 8.954735449735448,
|
|
"min": -9.735983397190292,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_full_share": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.005042559238173261,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.008374384236453203,
|
|
"min": -0.10829817158931077,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.12202178454078347,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.04132507369196753,
|
|
"repeat_abs_p95": 0.08139543508103343,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.12107279693486583,
|
|
"min": -0.0822222222222222,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_none_share": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.008375994890060249,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.09274496015002344,
|
|
"min": -0.01986863711001642,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.75,
|
|
"negative": 9,
|
|
"positive": 3,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.22398385207837773,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.037395530134597704,
|
|
"repeat_abs_p95": 0.07586812708364432,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.07631077509373656,
|
|
"min": -0.10941890166028097,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_padding_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0011703901376772582,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.006703036588479568,
|
|
"min": -0.0015820428864207082,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.75,
|
|
"negative": 3,
|
|
"positive": 9,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.29946486376659237,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.003908271985422232,
|
|
"repeat_abs_p95": 0.008187883039017584,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.007141591578301508,
|
|
"min": -0.009201187757062074,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_end_minus_start": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.006820888350695331,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.140893666867539,
|
|
"min": -0.005054735620165451,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.8333333333333334,
|
|
"negative": 2,
|
|
"positive": 10,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6661896437498689,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.010238658638254572,
|
|
"repeat_abs_p95": 0.1277225497052652,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.13402937356379263,
|
|
"min": -0.1168947946847837,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_max": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.007604372340016818,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.06684057734774362,
|
|
"min": 0.00016629145416469093,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.3770466591391387,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.020168252802926,
|
|
"repeat_abs_p95": 0.11084274153262068,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.12803476870816266,
|
|
"min": -0.11234888600259774,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_mean": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0014831270777658032,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.02895016189128262,
|
|
"min": -0.00029784868681948656,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 4,
|
|
"positive": 8,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.12120157645159224,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.01223686292857884,
|
|
"repeat_abs_p95": 0.07958153755957757,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.08820147130614941,
|
|
"min": -0.08208138515366167,
|
|
"n": 24
|
|
}
|
|
},
|
|
"preemptions": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0,
|
|
"action_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.0,
|
|
"negative": 0,
|
|
"positive": 0,
|
|
"zero": 12
|
|
},
|
|
"effect_to_repeat_median": 0.0,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0,
|
|
"repeat_abs_p95": 0.0,
|
|
"repeat_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 24
|
|
}
|
|
},
|
|
"prefill_token_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0028294297587345696,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.013691652669458043,
|
|
"min": -0.002949832419021914,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.44362661948582977,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.006377953067861264,
|
|
"repeat_abs_p95": 0.031272369907012425,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.036461564751038456,
|
|
"min": -0.018166035027340932,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_running_mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.50741128125,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 15.614046583666667,
|
|
"min": -0.30207624649999953,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.12962585804860988,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 3.9144294887499997,
|
|
"repeat_abs_p95": 7.101594725466666,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 6.513803891666667,
|
|
"min": -13.5536487565,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_waiting_mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.24895167858333334,
|
|
"action_delta": {
|
|
"distinct_n": 10,
|
|
"max": 0.0,
|
|
"min": -8.040727011166666,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 9,
|
|
"positive": 0,
|
|
"zero": 3
|
|
},
|
|
"effect_to_repeat_median": 1.1132189691456902,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.22363226416666668,
|
|
"repeat_abs_p95": 2.296808863758333,
|
|
"repeat_delta": {
|
|
"distinct_n": 16,
|
|
"max": 2.3311656403333334,
|
|
"min": -5.974843045,
|
|
"n": 24
|
|
}
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 3.249999999999993,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 10.5,
|
|
"min": -30.999999999999993,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.1065573770491801,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 30.5,
|
|
"repeat_abs_p95": 59.92499999999999,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 59.5,
|
|
"min": -64.5,
|
|
"n": 24
|
|
}
|
|
}
|
|
},
|
|
"sanity": {
|
|
"action_pairs": 12,
|
|
"invariants": {
|
|
"efficacy_label_balance": true,
|
|
"expected_action_pair_count": true,
|
|
"expected_repeat_pair_count": true,
|
|
"finite_deltas": true,
|
|
"matched_action_request_hashes": true,
|
|
"probabilities_bounded": true
|
|
},
|
|
"red_flags": [],
|
|
"repeat_pairs": 24,
|
|
"trials": 36
|
|
},
|
|
"start_s": 0.0
|
|
},
|
|
{
|
|
"action_pairs": 12,
|
|
"coverage_at_end": {
|
|
"admitted_fraction_of_total": {
|
|
"distinct_n": 22,
|
|
"max": 0.20535714285714285,
|
|
"median": 0.18235294117647058,
|
|
"min": 0.1564245810055866,
|
|
"n": 36
|
|
},
|
|
"completed_fraction_of_total": {
|
|
"distinct_n": 31,
|
|
"max": 0.180327868852459,
|
|
"median": 0.14411764705882352,
|
|
"min": 0.08933333333333335,
|
|
"n": 36
|
|
}
|
|
},
|
|
"efficacy": {
|
|
"feasibility_transitions": {
|
|
"false->false": 3,
|
|
"false->true": 6,
|
|
"true->true": 3
|
|
},
|
|
"minimum_balanced_accuracy": 0.75,
|
|
"minimum_delta_over_best_outcome": 0.15,
|
|
"outcome_delta": {
|
|
"best_accuracy": 0.8333333333333334,
|
|
"best_balanced_accuracy": 0.8333333333333333,
|
|
"best_feature": "ttft_max_over_slo_max",
|
|
"features": {
|
|
"admitted_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"admitted_input_tokens_mean_over_limit": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_fail_fraction_of_total": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.012666666666666666,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.008666666666666666,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.008666666666666666,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.03401360544217691,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.006172839506172811,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.04871948779511809,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_pass_rate": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.11875000000000002,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.07738095238095238,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.07738095238095238,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"outstanding_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.03401360544217688,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.006172839506172839,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.04871948779511806,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_max_over_slo": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.028440662518539606,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.02030446551608167,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.008882265987138677,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_mean_over_slo": {
|
|
"accuracy": 0.4166666666666667,
|
|
"balanced_accuracy": 0.41666666666666663,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.0020290170113650557,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.04992359808650447,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.009324194656287296,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_max_over_slo_max": {
|
|
"accuracy": 0.8333333333333334,
|
|
"balanced_accuracy": 0.8333333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.01753702182516767,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.01753702182516767,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.018086205842943556,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_mean_over_slo_max": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.010556909058063094,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.003084423316020539,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.027367186208561683,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_delta": {
|
|
"best_accuracy": 0.6666666666666666,
|
|
"best_balanced_accuracy": 0.6666666666666666,
|
|
"best_feature": "scheduler_steps_per_s",
|
|
"features": {
|
|
"decode_batch_size.mean": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.17161348298791945,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.30619421565755833,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.6654132739143297,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"graph_padding_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.005113976430465315,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -5.156781096974641e-05,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.0022397228564162543,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"kv_usage_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -1.7899149433659497e-05,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -2.0776132762541973e-05,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00013983916927168966,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"prefill_token_fraction": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 3.837762108915843e-05,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.00021675516061736566,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 3.837762108915843e-05,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_running_mean": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.0830968510000003,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.12225680983333342,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.2734091407916659,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_waiting_mean": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.2763032289583333,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.08217400995833332,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -1.0689820661666667,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 3.2916666666666643,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 3.5833333333333286,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 6.291666666666657,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_qualifying_features": []
|
|
},
|
|
"end_fraction": 0.2,
|
|
"end_s": 12.0,
|
|
"qualifying_response_features": [],
|
|
"repeat_pairs": 24,
|
|
"response_statistics": {
|
|
"batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.6805455173736323,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 21.022824302134648,
|
|
"min": -1.6199557480463245,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.2914327913831549,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 2.335171392840622,
|
|
"repeat_abs_p95": 10.342426400114645,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 10.582610461920808,
|
|
"min": -14.50489117143556,
|
|
"n": 24
|
|
}
|
|
},
|
|
"batch_tokens.mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 30.73978867286374,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 837.5693088520675,
|
|
"min": -38.259099896431565,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 4,
|
|
"positive": 8,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.5085540691379722,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 60.44546792235762,
|
|
"repeat_abs_p95": 370.0653796793133,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 385.95091643367505,
|
|
"min": -701.2529783740349,
|
|
"n": 24
|
|
}
|
|
},
|
|
"decode_batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.6733471921553615,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 20.720637408568443,
|
|
"min": -1.606656623670089,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.2892646285380025,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 2.327789593766041,
|
|
"repeat_abs_p95": 10.226231015894982,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 10.460062847993882,
|
|
"min": -14.289496209342136,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_full_share": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.008498282584807226,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.024842293569343754,
|
|
"min": -0.15582922824302126,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 7,
|
|
"positive": 5,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.517404626039514,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.01642482915133081,
|
|
"repeat_abs_p95": 0.0799421671518786,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.11509974496034647,
|
|
"min": -0.08092575333954644,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_none_share": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.01137367331008028,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.14248395282878046,
|
|
"min": -0.030463233217211186,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 8,
|
|
"positive": 4,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6710383267094245,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.016949364674672823,
|
|
"repeat_abs_p95": 0.07611571753021439,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.0761583865032141,
|
|
"min": -0.11305593403905953,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_padding_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.0022211647342633674,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.007755344475118115,
|
|
"min": -0.0006496983398026016,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.8333333333333334,
|
|
"negative": 2,
|
|
"positive": 10,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.0130418206479903,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0021925696343341517,
|
|
"repeat_abs_p95": 0.006616310818721587,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.007943398472261458,
|
|
"min": -0.005259691066565701,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_end_minus_start": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.0021058619896323627,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.06906439560383515,
|
|
"min": -0.0026342182586551743,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.75,
|
|
"negative": 3,
|
|
"positive": 9,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.17774345954986717,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.011847760783802841,
|
|
"repeat_abs_p95": 0.0656650541454329,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.020281746428214564,
|
|
"min": -0.10385652912378862,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_max": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.011216828476763308,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.08922043547542868,
|
|
"min": 0.0002683613903166071,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.0844557892569866,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.010343278709820436,
|
|
"repeat_abs_p95": 0.07465434288403082,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.05015486062543706,
|
|
"min": -0.12688580277750028,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_mean": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0041456974144474384,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.047957700246711504,
|
|
"min": -0.00037567596362376825,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.75,
|
|
"negative": 3,
|
|
"positive": 9,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.8378282569881378,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.004948147045494235,
|
|
"repeat_abs_p95": 0.06813512586166338,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.07105816994418354,
|
|
"min": -0.09865864218773521,
|
|
"n": 24
|
|
}
|
|
},
|
|
"preemptions": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0,
|
|
"action_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.0,
|
|
"negative": 0,
|
|
"positive": 0,
|
|
"zero": 12
|
|
},
|
|
"effect_to_repeat_median": 0.0,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0,
|
|
"repeat_abs_p95": 0.0,
|
|
"repeat_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 24
|
|
}
|
|
},
|
|
"prefill_token_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0003854997463232923,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.012211161699714035,
|
|
"min": -0.0029429188195312372,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.04836213965959768,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.007971106097386826,
|
|
"repeat_abs_p95": 0.020160934134987728,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.011632967170489006,
|
|
"min": -0.023705586156158254,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_running_mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.5689573993333328,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 27.59238343825,
|
|
"min": -0.6469868719999994,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.26512988706199775,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 2.1459572349166667,
|
|
"repeat_abs_p95": 7.856052630799997,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 6.557133349166667,
|
|
"min": -17.1293622325,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_waiting_mean": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 0.40246922141666663,
|
|
"action_delta": {
|
|
"distinct_n": 11,
|
|
"max": 0.0,
|
|
"min": -19.377070914,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 10,
|
|
"positive": 0,
|
|
"zero": 2
|
|
},
|
|
"effect_to_repeat_median": 4.226054832868273,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.09523521045833332,
|
|
"repeat_abs_p95": 2.510103133283333,
|
|
"repeat_delta": {
|
|
"distinct_n": 18,
|
|
"max": 2.268429387666668,
|
|
"min": -11.883906298916667,
|
|
"n": 24
|
|
}
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 7.958333333333329,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 11.75,
|
|
"min": -35.16666666666667,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 7,
|
|
"positive": 5,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.38047808764940205,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 20.91666666666667,
|
|
"repeat_abs_p95": 51.31666666666662,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 65.66666666666667,
|
|
"min": -34.58333333333333,
|
|
"n": 24
|
|
}
|
|
}
|
|
},
|
|
"sanity": {
|
|
"action_pairs": 12,
|
|
"invariants": {
|
|
"efficacy_label_balance": true,
|
|
"expected_action_pair_count": true,
|
|
"expected_repeat_pair_count": true,
|
|
"finite_deltas": true,
|
|
"matched_action_request_hashes": true,
|
|
"probabilities_bounded": true
|
|
},
|
|
"red_flags": [],
|
|
"repeat_pairs": 24,
|
|
"trials": 36
|
|
},
|
|
"start_s": 0.0
|
|
},
|
|
{
|
|
"action_pairs": 12,
|
|
"coverage_at_end": {
|
|
"admitted_fraction_of_total": {
|
|
"distinct_n": 22,
|
|
"max": 0.30654761904761907,
|
|
"median": 0.2795403407264545,
|
|
"min": 0.24313725490196078,
|
|
"n": 36
|
|
},
|
|
"completed_fraction_of_total": {
|
|
"distinct_n": 30,
|
|
"max": 0.2681564245810056,
|
|
"median": 0.22518087736972248,
|
|
"min": 0.12849162011173185,
|
|
"n": 36
|
|
}
|
|
},
|
|
"efficacy": {
|
|
"feasibility_transitions": {
|
|
"false->false": 3,
|
|
"false->true": 6,
|
|
"true->true": 3
|
|
},
|
|
"minimum_balanced_accuracy": 0.75,
|
|
"minimum_delta_over_best_outcome": 0.15,
|
|
"outcome_delta": {
|
|
"best_accuracy": 0.75,
|
|
"best_balanced_accuracy": 0.75,
|
|
"best_feature": "completed_pass_rate",
|
|
"features": {
|
|
"admitted_fraction": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"admitted_input_tokens_mean_over_limit": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0,
|
|
"train_balanced_accuracy": 0.5
|
|
}
|
|
]
|
|
},
|
|
"completed_fail_fraction_of_total": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.02825325884543762,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.02266666666666667,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.040253258845437616,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.03272660415517559,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.022522522522522515,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.07358436332279389,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"completed_pass_rate": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.14270516717325227,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.12142857142857144,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.2182462927143778,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"outstanding_over_admitted": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.03272660415517557,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.022522522522522515,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.0735843633227939,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"tpot_max_over_slo": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 0.37412821213021774,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.009857117327656123,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.008882265987138677,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"tpot_mean_over_slo": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666667,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.004552157689988692,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.007167514027906635,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.006851562902858294,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_max_over_slo_max": {
|
|
"accuracy": 0.5,
|
|
"balanced_accuracy": 0.5,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": -0.017755539496041216,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.007733676747496546,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.5040019049192779,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"ttft_mean_over_slo_max": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.007216190772090309,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.007216190772090309,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.24525856553394837,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_delta": {
|
|
"best_accuracy": 0.75,
|
|
"best_balanced_accuracy": 0.75,
|
|
"best_feature": "scheduler_steps_per_s",
|
|
"features": {
|
|
"decode_batch_size.mean": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.44507332514357945,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.24667983170250252,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.5943046676188155,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"graph_padding_fraction": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333334,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.004225031845271454,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0004360066659713927,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0034270719814960106,
|
|
"train_balanced_accuracy": 0.875
|
|
}
|
|
]
|
|
},
|
|
"kv_usage_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.00013357357159209972,
|
|
"train_balanced_accuracy": 0.625
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.00013492200088597965,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": -0.00018506887864057613,
|
|
"train_balanced_accuracy": 0.625
|
|
}
|
|
]
|
|
},
|
|
"prefill_token_fraction": {
|
|
"accuracy": 0.5833333333333334,
|
|
"balanced_accuracy": 0.5833333333333333,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 0.0008649236257326498,
|
|
"train_balanced_accuracy": 1.0
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": 0.0017369622931825424,
|
|
"train_balanced_accuracy": 0.875
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": 0.0005824031404205732,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_running_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 11.957967036527776,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
0
|
|
],
|
|
"threshold": -0.23078758088888907,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
1,
|
|
0
|
|
],
|
|
"threshold": 18.57988346463889,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"queue_waiting_mean": {
|
|
"accuracy": 0.6666666666666666,
|
|
"balanced_accuracy": 0.6666666666666666,
|
|
"folds": [
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.18938916494444447,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
1,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.12479013208333334,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": -1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
1,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"threshold": -0.23008102869444447,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"accuracy": 0.75,
|
|
"balanced_accuracy": 0.75,
|
|
"folds": [
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 1,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 4.277777777777775,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 2,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 3.888888888888893,
|
|
"train_balanced_accuracy": 0.75
|
|
},
|
|
{
|
|
"direction": 1,
|
|
"held_out_replicate": 3,
|
|
"test_labels": [
|
|
0,
|
|
0,
|
|
1,
|
|
1
|
|
],
|
|
"test_predictions": [
|
|
0,
|
|
0,
|
|
0,
|
|
1
|
|
],
|
|
"threshold": 4.277777777777775,
|
|
"train_balanced_accuracy": 0.75
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"labels": {
|
|
"distinct_n": 2,
|
|
"max": 1.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"negative": 6,
|
|
"positive": 6
|
|
},
|
|
"telemetry_qualifying_features": []
|
|
},
|
|
"end_fraction": 0.3,
|
|
"end_s": 18.0,
|
|
"qualifying_response_features": [],
|
|
"repeat_pairs": 24,
|
|
"response_statistics": {
|
|
"batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.6066330085464959,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 28.73606563075747,
|
|
"min": -1.1485291194593508,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.4934987464973646,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 1.2292493402508278,
|
|
"repeat_abs_p95": 8.067582719584564,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 8.686808262759335,
|
|
"min": -14.918642465494425,
|
|
"n": 24
|
|
}
|
|
},
|
|
"batch_tokens.mean": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 10.317307738579501,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 801.7642402753155,
|
|
"min": -13.613525901316592,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.17112975179373358,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 60.28938644763055,
|
|
"repeat_abs_p95": 340.3503302119946,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 374.64907917730886,
|
|
"min": -561.6272830402723,
|
|
"n": 24
|
|
}
|
|
},
|
|
"decode_batch_size.mean": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.6035604856985604,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 28.411725240727222,
|
|
"min": -1.1421295113155576,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.4962266942146829,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 1.2162999144045272,
|
|
"repeat_abs_p95": 7.9623825938693935,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 8.570025092978447,
|
|
"min": -14.737105312913595,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_full_share": {
|
|
"action_above_repeat_p95_fraction": 0.25,
|
|
"action_abs_median": 0.006690006870484877,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.013223015997172571,
|
|
"min": -0.1490562102409011,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5833333333333334,
|
|
"negative": 5,
|
|
"positive": 7,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.6364709072202164,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.010511096099747042,
|
|
"repeat_abs_p95": 0.06858287350501925,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.08329551900170162,
|
|
"min": -0.0738002419680065,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_none_share": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.01055037975803224,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.13907776271422118,
|
|
"min": -0.018167596600093858,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 8,
|
|
"positive": 4,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.9080980815382388,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.01161810598714273,
|
|
"repeat_abs_p95": 0.06969738634332051,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.07468521754716137,
|
|
"min": -0.08558328606541879,
|
|
"n": 24
|
|
}
|
|
},
|
|
"graph_padding_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.16666666666666666,
|
|
"action_abs_median": 0.0024714849986972203,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.006494393749041705,
|
|
"min": 0.000227090832439927,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.373420385875363,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0017995109320603211,
|
|
"repeat_abs_p95": 0.005547351274885883,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.005694038006766439,
|
|
"min": -0.0035699364880986415,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_end_minus_start": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.010270226283016048,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.23522057920397632,
|
|
"min": -0.003022494939389775,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.9166666666666666,
|
|
"negative": 1,
|
|
"positive": 11,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.07628737944467,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.009542271403679525,
|
|
"repeat_abs_p95": 0.08564791687481275,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.08662204016385266,
|
|
"min": -0.31416724947547214,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_max": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.015846588112988946,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.22224967428314468,
|
|
"min": 0.0002683613903166071,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 0,
|
|
"positive": 12,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 1.3242319261149558,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.011966625936500275,
|
|
"repeat_abs_p95": 0.10502297931861322,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.05015486062543706,
|
|
"min": -0.14511939254670791,
|
|
"n": 24
|
|
}
|
|
},
|
|
"kv_usage_mean": {
|
|
"action_above_repeat_p95_fraction": 0.08333333333333333,
|
|
"action_abs_median": 0.004416664987538066,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.06671902028938567,
|
|
"min": -0.00038309683389285525,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.8333333333333334,
|
|
"negative": 2,
|
|
"positive": 10,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.8499259269415633,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.005196529306302399,
|
|
"repeat_abs_p95": 0.03531051450496546,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.028898615902990022,
|
|
"min": -0.05744492686343953,
|
|
"n": 24
|
|
}
|
|
},
|
|
"preemptions": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.0,
|
|
"action_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.0,
|
|
"negative": 0,
|
|
"positive": 0,
|
|
"zero": 12
|
|
},
|
|
"effect_to_repeat_median": 0.0,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": false,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.0,
|
|
"repeat_abs_p95": 0.0,
|
|
"repeat_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"min": 0.0,
|
|
"n": 24
|
|
}
|
|
},
|
|
"prefill_token_fraction": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 0.001539018865967745,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 0.003050797976300035,
|
|
"min": -0.0015887887661509836,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 4,
|
|
"positive": 8,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.15591584638446954,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.009870830333516656,
|
|
"repeat_abs_p95": 0.021489430916148855,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 0.021795755533912664,
|
|
"min": -0.02021322592113628,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_running_mean": {
|
|
"action_above_repeat_p95_fraction": 0.3333333333333333,
|
|
"action_abs_median": 0.5940786813055561,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 33.446711043166665,
|
|
"min": -0.6256704651111118,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.6666666666666666,
|
|
"negative": 4,
|
|
"positive": 8,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.30447578852990603,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 1.9511524518055556,
|
|
"repeat_abs_p95": 4.689088691175,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 3.8720077572777782,
|
|
"min": -11.837926215166668,
|
|
"n": 24
|
|
}
|
|
},
|
|
"queue_waiting_mean": {
|
|
"action_above_repeat_p95_fraction": 0.4166666666666667,
|
|
"action_abs_median": 0.9777240376944445,
|
|
"action_delta": {
|
|
"distinct_n": 11,
|
|
"max": 0.0,
|
|
"min": -28.609993227777778,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 1.0,
|
|
"negative": 10,
|
|
"positive": 0,
|
|
"zero": 2
|
|
},
|
|
"effect_to_repeat_median": 8.318675069172114,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 0.11753362519444445,
|
|
"repeat_abs_p95": 2.0560520443777768,
|
|
"repeat_delta": {
|
|
"distinct_n": 20,
|
|
"max": 1.6669221998333335,
|
|
"min": -9.630657473833335,
|
|
"n": 24
|
|
}
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"action_above_repeat_p95_fraction": 0.0,
|
|
"action_abs_median": 5.527777777777779,
|
|
"action_delta": {
|
|
"distinct_n": 12,
|
|
"max": 9.166666666666671,
|
|
"min": -33.27777777777777,
|
|
"n": 12
|
|
},
|
|
"action_signs": {
|
|
"consistency": 0.5,
|
|
"negative": 6,
|
|
"positive": 6,
|
|
"zero": 0
|
|
},
|
|
"effect_to_repeat_median": 0.3338926174496646,
|
|
"effect_to_repeat_median_is_infinite": false,
|
|
"gate_feature": true,
|
|
"qualifies": false,
|
|
"repeat_abs_median": 16.55555555555555,
|
|
"repeat_abs_p95": 37.27222222222221,
|
|
"repeat_delta": {
|
|
"distinct_n": 24,
|
|
"max": 31.888888888888886,
|
|
"min": -41.611111111111114,
|
|
"n": 24
|
|
}
|
|
}
|
|
},
|
|
"sanity": {
|
|
"action_pairs": 12,
|
|
"invariants": {
|
|
"efficacy_label_balance": true,
|
|
"expected_action_pair_count": true,
|
|
"expected_repeat_pair_count": true,
|
|
"finite_deltas": true,
|
|
"matched_action_request_hashes": true,
|
|
"probabilities_bounded": true
|
|
},
|
|
"red_flags": [],
|
|
"repeat_pairs": 24,
|
|
"trials": 36
|
|
},
|
|
"start_s": 0.0
|
|
}
|
|
],
|
|
"decision": "REQUIRES_UNCENSORED_PHASE_AWARE_PILOT",
|
|
"design": {
|
|
"available_deciles": [
|
|
0.1,
|
|
0.2,
|
|
0.3
|
|
],
|
|
"cumulative_and_nonoverlapping_blocks": true,
|
|
"decile_fraction": 0.1,
|
|
"maximum_common_end_s": 18.0,
|
|
"maximum_common_fraction": 0.3,
|
|
"select_best_horizon": false,
|
|
"trace_duration_s": 60.0
|
|
},
|
|
"provenance": {
|
|
"analysis_script": "/home/gahow/phd/aituner/runs/intervention-response-v2/analyze_existing.py",
|
|
"analysis_script_sha256": "c7eb2cd2d00f2b3afa62e22e7b2f1cce6b87ec51ebb7921ef7ffb9d13d28c50a",
|
|
"manifest": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/pilot-manifest.json",
|
|
"manifest_sha256": "86b249a087ab1bcd51cd8de3d1da6bf34afd41af977eb58c93229b941dbf5c8a",
|
|
"manifest_validation": {
|
|
"expected_trials": 36,
|
|
"matched_trials": 36,
|
|
"schema": "fidelity-prefix-pilot-manifest-v1"
|
|
},
|
|
"p1_analysis_script": "/home/gahow/phd/aituner/runs/intervention-response-v0/analyze_p1.py",
|
|
"p1_analysis_script_sha256": "902adcba58bcb73f97c455fdea3710f6cb28fa8d8f4ea5bc4915143ed106b17a",
|
|
"run_root": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b",
|
|
"streams": [
|
|
{
|
|
"bytes": 24501686,
|
|
"cell": "tp1_mns64",
|
|
"path": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/cells/tp1_mns64/opprof/opprof-v1-dp0-pid234511-1784005631788937266.jsonl",
|
|
"sha256": "a813076afe677e5c13010b2ec44f46998794ca44cdf30ee81be836e3623f0068"
|
|
},
|
|
{
|
|
"bytes": 25524027,
|
|
"cell": "tp1_mns8",
|
|
"path": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/cells/tp1_mns8/opprof/opprof-v1-dp0-pid227734-1784005177936445360.jsonl",
|
|
"sha256": "1708759d2e77c42bff1940d43214fb2d927c3eebeda62d53a3fb7d04817e9b22"
|
|
},
|
|
{
|
|
"bytes": 29431988,
|
|
"cell": "tp2_mns64",
|
|
"path": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/cells/tp2_mns64/opprof/opprof-v1-dp0-pid247791-1784006495467408663.jsonl",
|
|
"sha256": "65355de883f2eb5d31be161d82b26dddc53c41a0a35a0fdb807777bc5b21c5ec"
|
|
},
|
|
{
|
|
"bytes": 26036993,
|
|
"cell": "tp2_mns8",
|
|
"path": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/cells/tp2_mns8/opprof/opprof-v1-dp0-pid241594-1784006096655191299.jsonl",
|
|
"sha256": "485e203d37d45fdd3c4de768b95da2050478f111f4d739b129d4f4ebd4737b70"
|
|
},
|
|
{
|
|
"bytes": 17449143,
|
|
"cell": "tp4_mns16",
|
|
"path": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/cells/tp4_mns16/opprof/opprof-v1-dp0-pid256090-1784007028237607411.jsonl",
|
|
"sha256": "59e98014d01bfc68d9f97112ecd782cb03571ad437382ed67416c7b85ae48c56"
|
|
},
|
|
{
|
|
"bytes": 22674678,
|
|
"cell": "tp4_mns64",
|
|
"path": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/cells/tp4_mns64/opprof/opprof-v1-dp0-pid262293-1784007376383569515.jsonl",
|
|
"sha256": "03e4f3cca3023a78e859b663703b9468dedf89e58320f0292e2973070e7616c8"
|
|
}
|
|
],
|
|
"trial_inputs": [
|
|
{
|
|
"cell": "tp1_mns64",
|
|
"early_stopped": true,
|
|
"elapsed_s": 19.448401608,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 1,
|
|
"request_count": 179,
|
|
"requests_sha256": "e970dcac20144cb1e13a35997377f6695a9c57d788a88bad7f18cf6c64e35cec",
|
|
"result_sha256": "3233686b60b920dcd6cc2f45193a81f3af2d46f781ef42c4d117630612084feb",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns64/high-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns64",
|
|
"early_stopped": true,
|
|
"elapsed_s": 56.362951594,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 2,
|
|
"request_count": 179,
|
|
"requests_sha256": "ec06878e2aefdc4853d972197cedc29ab94df7147abb2701e0d88b5dabf29728",
|
|
"result_sha256": "8215477a014f061c55e94d879a25bbf1676ee2ef1c5e81efadd23d20d337e07a",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns64/high-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns64",
|
|
"early_stopped": true,
|
|
"elapsed_s": 59.320339642,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 3,
|
|
"request_count": 179,
|
|
"requests_sha256": "18d72d1fe16136359e28a1909e395ffd553dbe3edd2c0fd93a9cc270a5ec6d60",
|
|
"result_sha256": "480a3b50acaf9b9c1902e2669cf3aa7723a6703a00778da269b5b9757990ceef",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns64/high-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.653813767,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 1,
|
|
"request_count": 122,
|
|
"requests_sha256": "44d729c1ace9b09fe2ea4c4e110bd0f95fee69f83edc78ee0f65256922dcd2c2",
|
|
"result_sha256": "d2c08e71c5d4e0dd99662bab8457bd1ad0a4aba6bda57794dbfce62e36720f62",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns64/low-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.821918535,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 2,
|
|
"request_count": 122,
|
|
"requests_sha256": "97de3cced64fd88a5c2dcd087ca6c29d8b09a0a43d7bf0e4e844ef246462465b",
|
|
"result_sha256": "f194db2f7abb995171cd24626e4e515c000481f81283c929b9a5b63f110f1506",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns64/low-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 59.372640014,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 3,
|
|
"request_count": 122,
|
|
"requests_sha256": "30f4623100c58580cfa4efe9a1998c63581c8e44511c556ef39cf5bd777883bf",
|
|
"result_sha256": "0be457aa2fc9ce95bdb92a6eb25544f5f61a9739932ee5afa40535d26456e348",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns64/low-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns8",
|
|
"early_stopped": true,
|
|
"elapsed_s": 24.256301395,
|
|
"level": "high",
|
|
"mns": 8,
|
|
"replicate": 1,
|
|
"request_count": 179,
|
|
"requests_sha256": "5f5bfa4a16846190f488d3369eed89ae895804145cea59197e6b544b703223fc",
|
|
"result_sha256": "f65e3dc91d17f5caab1f4aecf97d2c464b99f107231af2cfd96f5484880136ee",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns8/high-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns8",
|
|
"early_stopped": true,
|
|
"elapsed_s": 57.560632316,
|
|
"level": "high",
|
|
"mns": 8,
|
|
"replicate": 2,
|
|
"request_count": 179,
|
|
"requests_sha256": "e802d09057de349e5164c18314fcad09dc88c9673834fd6e5b100a57dd12fbf4",
|
|
"result_sha256": "4835a985dd6bb8dc50b53737baa980ac95a49e22f0b8a7326e1fb416c9b4e798",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns8/high-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns8",
|
|
"early_stopped": true,
|
|
"elapsed_s": 46.401123902,
|
|
"level": "high",
|
|
"mns": 8,
|
|
"replicate": 3,
|
|
"request_count": 179,
|
|
"requests_sha256": "cf1226185845cd8aa4e7ead0cb5c95c77eda62b84851a2ab579f984bbb88e8ac",
|
|
"result_sha256": "95d5103f2678eeb99c258870324566932e2f2853c3b6e0ee94bc6c538f6ed865",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns8/high-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns8",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.653530104,
|
|
"level": "low",
|
|
"mns": 8,
|
|
"replicate": 1,
|
|
"request_count": 122,
|
|
"requests_sha256": "0edc6095b16e5361a88b2765cc5d0f267ed16ff1d883e02fc3c58f57e59c1018",
|
|
"result_sha256": "8bc433109d148c21b7d158df95ec5d1fa4fa619243114563068e1e98c4241a29",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns8/low-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns8",
|
|
"early_stopped": false,
|
|
"elapsed_s": 62.156958699,
|
|
"level": "low",
|
|
"mns": 8,
|
|
"replicate": 2,
|
|
"request_count": 122,
|
|
"requests_sha256": "ca58a878401ac44cca0a9a595eab41659dafec2e923f7f2474e4f82ea30b2f3e",
|
|
"result_sha256": "42a56af136d1d3bdfc16d9cea092c2db0eaa4db2c85dceb21958c5e418122900",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns8/low-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp1_mns8",
|
|
"early_stopped": false,
|
|
"elapsed_s": 59.499655201,
|
|
"level": "low",
|
|
"mns": 8,
|
|
"replicate": 3,
|
|
"request_count": 122,
|
|
"requests_sha256": "8a72240f84bbeff9eabd63e7d1f7a4c19910660fc5e4335013a1deaf89408127",
|
|
"result_sha256": "0214f32cabbcdc1bad6c331c3945c505cad136fcc9a8336aeb6314b27474e279",
|
|
"tp": 1,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp1_mns8/low-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.842664607,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 1,
|
|
"request_count": 345,
|
|
"requests_sha256": "19537ec05f1e40543d850964db12234561954c209d55cd8cd0a1da716ab105fc",
|
|
"result_sha256": "49523b3269db649aa70072058666da3309ecabb93242cd504b6413f5de0f5d54",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns64/high-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.33597184,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 2,
|
|
"request_count": 345,
|
|
"requests_sha256": "d95e0dcc3d3f25af11b4d3da0ca133f647fd602753b804337a7131cc56e03059",
|
|
"result_sha256": "f37a42d9ea734adcb24892f86cbb19e2c51cee0baf8a6a63c5d51c4820bd28f2",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns64/high-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.925907661,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 3,
|
|
"request_count": 345,
|
|
"requests_sha256": "ccba3d0a18a7db8501b040855bda473ee55c519619d4b9fa2a105bda06300829",
|
|
"result_sha256": "55447419925f30397d19863e149890241c58863d0adaafd511a67757905cb58e",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns64/high-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.73903317,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 1,
|
|
"request_count": 235,
|
|
"requests_sha256": "8685f475baf247f7ce52a6a2df1183cd77a6f580cd57a3a87c407b4e96c0895f",
|
|
"result_sha256": "7a8f4e462d496bd56b73b0a686653cbe499a238df62a8c9a50db687d051f1c68",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns64/low-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.049384964,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 2,
|
|
"request_count": 235,
|
|
"requests_sha256": "016a76ab3dc015700f9ecb9f60a20fcc3b456c45270472fdd1d04e84ebd7021a",
|
|
"result_sha256": "368869326f44a0032985eff9d3abfa2306b93c25370caa05cee255a8c0208c0e",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns64/low-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.611333104,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 3,
|
|
"request_count": 235,
|
|
"requests_sha256": "5455d1cdeabd6e2937130ed8fc8d86c1aaa6f04cbf26c3cb69ad064be5f67d73",
|
|
"result_sha256": "2de5c4524f57b487b94b4fdba35defc2a566236fc3e31f7260b7a9f85bb7007c",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns64/low-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns8",
|
|
"early_stopped": true,
|
|
"elapsed_s": 25.452006333,
|
|
"level": "high",
|
|
"mns": 8,
|
|
"replicate": 1,
|
|
"request_count": 336,
|
|
"requests_sha256": "5d41a316f329e2efcc0f728cdd078377eb72fb9258738636db704bcfcca290aa",
|
|
"result_sha256": "466b3991190a232c984ba1b54fb6793cac1893dce196ee96f7238d70d49fcb74",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns8/high-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns8",
|
|
"early_stopped": true,
|
|
"elapsed_s": 24.770382696,
|
|
"level": "high",
|
|
"mns": 8,
|
|
"replicate": 2,
|
|
"request_count": 336,
|
|
"requests_sha256": "1b7d2cdcd639c00c443ae1b98287a7cc57323eb80c7f0e1bfc72f1ab5bf48760",
|
|
"result_sha256": "e9156d6871cc8cb3a20dce629cc2443df7473f3836b46eb0944d491ea5c4d44c",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns8/high-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns8",
|
|
"early_stopped": true,
|
|
"elapsed_s": 25.69871505,
|
|
"level": "high",
|
|
"mns": 8,
|
|
"replicate": 3,
|
|
"request_count": 336,
|
|
"requests_sha256": "5a68fda7c29477e5b3e66d70ad2b05b77be6c200cc1beea89d663a5729f31a2b",
|
|
"result_sha256": "f2116797b35057949acbf1f57898befac9ff18b94c808462f0969bdcf2446dd6",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns8/high-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns8",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.605998599,
|
|
"level": "low",
|
|
"mns": 8,
|
|
"replicate": 1,
|
|
"request_count": 229,
|
|
"requests_sha256": "455a1c21388700f0e3234c49d15ba3b0426b0d3b551c096b297793afa0d4602f",
|
|
"result_sha256": "13de25a73cc8c752245f169cfca1d64220f697c15b835a646ba969a1005047fb",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns8/low-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns8",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.925251312,
|
|
"level": "low",
|
|
"mns": 8,
|
|
"replicate": 2,
|
|
"request_count": 229,
|
|
"requests_sha256": "cae4b764ebb8714a6f5d2009bb79497420b388986a992c51410f1681810c14c6",
|
|
"result_sha256": "92f92a576e990946c4767a2107d2ce0e6ba1b066337911399360c199fae3f5e4",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns8/low-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp2_mns8",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.612475588,
|
|
"level": "low",
|
|
"mns": 8,
|
|
"replicate": 3,
|
|
"request_count": 229,
|
|
"requests_sha256": "cc8c01e4560411cd5b84cdc9bebdda9a86b88615ee1c39a996affe669d26c167",
|
|
"result_sha256": "c9a7d026bd89cc027c45b8f1899d56e21d53d4234e95db365a9492f670a78d36",
|
|
"tp": 2,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp2_mns8/low-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns16",
|
|
"early_stopped": true,
|
|
"elapsed_s": 25.468414615,
|
|
"level": "high",
|
|
"mns": 16,
|
|
"replicate": 1,
|
|
"request_count": 750,
|
|
"requests_sha256": "8695422a0f1e0e1c097a08a7f4c082f1ba2e3bb26b58caadae0afc7438aeca90",
|
|
"result_sha256": "c54a2bb5c1b78bdef3fa09c39f5974215370d6fc7c2b6d403a3d376d53a1d9eb",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns16/high-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns16",
|
|
"early_stopped": true,
|
|
"elapsed_s": 24.508063881,
|
|
"level": "high",
|
|
"mns": 16,
|
|
"replicate": 2,
|
|
"request_count": 750,
|
|
"requests_sha256": "aa0482b10414d911af227e2df172c19ecca0add2e0024dc88f85ffbd063854bb",
|
|
"result_sha256": "91e2491de4ca0868b667e7d126b755f03bdd78378681b8070a2bbc0c38fb9055",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns16/high-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns16",
|
|
"early_stopped": true,
|
|
"elapsed_s": 27.456202024,
|
|
"level": "high",
|
|
"mns": 16,
|
|
"replicate": 3,
|
|
"request_count": 750,
|
|
"requests_sha256": "3a23ffbe0e724648226152ef5433cc5825ad640c37cf47997c2f091210eceeec",
|
|
"result_sha256": "136dde0451f806841ff3996b081fe0bffac0a76f60c3d78b7516f2689e106008",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns16/high-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns16",
|
|
"early_stopped": true,
|
|
"elapsed_s": 48.952560958,
|
|
"level": "low",
|
|
"mns": 16,
|
|
"replicate": 1,
|
|
"request_count": 510,
|
|
"requests_sha256": "d7eaaefd35cc906f759dcaa7ac3c8543249a1a02f28572bec953e41e4379a29f",
|
|
"result_sha256": "11d9ec9d2a3358aed988f7b4b1014af087d61c4b5a72eb9076ae19bb5cc481cd",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns16/low-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns16",
|
|
"early_stopped": true,
|
|
"elapsed_s": 49.912113702,
|
|
"level": "low",
|
|
"mns": 16,
|
|
"replicate": 2,
|
|
"request_count": 510,
|
|
"requests_sha256": "7fdb90f679bba1747fbacc92fac2045c820e09965c755b20f6dafd4fd9e26a7a",
|
|
"result_sha256": "800e2aa6e9f9efef4be12ee778be6817364914d63ae0624775d02729089768d6",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns16/low-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns16",
|
|
"early_stopped": true,
|
|
"elapsed_s": 48.923086251,
|
|
"level": "low",
|
|
"mns": 16,
|
|
"replicate": 3,
|
|
"request_count": 510,
|
|
"requests_sha256": "b592fe254476957265fc9139aefdea77be4b5540f3c1b8a429c10ea6ccab9a6a",
|
|
"result_sha256": "cc1b1f7f5280c2892d0c6deac9df9f34ce126b0284e6ad893ffbf0ddbf27969f",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns16/low-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.434823743,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 1,
|
|
"request_count": 750,
|
|
"requests_sha256": "b4aeaf5b3175bd85fdac08a2ef657fb2278c6c4b36ea32ea417275ab6d55812d",
|
|
"result_sha256": "111a1042b614d05772888665a7cfbc86925b2fbeb52ffe19edb62a5bcfe2f513",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns64/high-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.130504059,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 2,
|
|
"request_count": 750,
|
|
"requests_sha256": "5d82b6885f5c654ae1ae126696df4bafe04f5be1c1109db24c7856b0456d9443",
|
|
"result_sha256": "6ed574c18723e265ab1327c5e43d3c76d37b3fc5711d06c77a51718e4ba16601",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns64/high-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.22613867,
|
|
"level": "high",
|
|
"mns": 64,
|
|
"replicate": 3,
|
|
"request_count": 750,
|
|
"requests_sha256": "5c5861d716181f06dc3f951ae0d6ee479b1b61fc21ebf24541c5b45053fcc167",
|
|
"result_sha256": "35b8f0465f6b0c197f71b5ecaf4e771e6613f86998ae44f8701ab9e4e1e440b4",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns64/high-rep3/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 61.123849552,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 1,
|
|
"request_count": 510,
|
|
"requests_sha256": "931d372ded1e3aaa9c67f06653fac8f79b4ee4f369028c1148163580386681ce",
|
|
"result_sha256": "325ce57f33133670a6761b5dffb831ed3b2f464755ff743b56976b4a974c1215",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns64/low-rep1/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.86041954,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 2,
|
|
"request_count": 510,
|
|
"requests_sha256": "72c8f70a240f79edcaa4eb0a2fe9bade8ac07093339c5b2d09c49ddd1cf27d16",
|
|
"result_sha256": "ff8b2283d02d5967893ca6225b561ae11011da6149d1f6cff7b4275c00170e1d",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns64/low-rep2/result.json"
|
|
},
|
|
{
|
|
"cell": "tp4_mns64",
|
|
"early_stopped": false,
|
|
"elapsed_s": 60.47213624,
|
|
"level": "low",
|
|
"mns": 64,
|
|
"replicate": 3,
|
|
"request_count": 510,
|
|
"requests_sha256": "8b2c779d2097c5e4038e94e1ac9b986e29f55d9acfd67eeb366c4e0e9270f6e0",
|
|
"result_sha256": "f026b69356701894fab184e0f2f7504a30b537982e823373899c9ab123a33891",
|
|
"tp": 4,
|
|
"trace_duration_s": 60.0,
|
|
"trial_id": "cells/tp4_mns64/low-rep3/result.json"
|
|
}
|
|
]
|
|
},
|
|
"sanity": {
|
|
"early_stopped": 15,
|
|
"elapsed_s": {
|
|
"distinct_n": 36,
|
|
"max": 62.156958699,
|
|
"median": 60.5390674195,
|
|
"min": 19.448401608,
|
|
"n": 36
|
|
},
|
|
"invariants": {
|
|
"all_intervals_uncensored": true,
|
|
"all_window_sanity_pass": true,
|
|
"expected_trial_count": true,
|
|
"manifest_trials_match": true,
|
|
"stream_provenance_consistent": true,
|
|
"trace_duration_consistent": true
|
|
},
|
|
"red_flags": [],
|
|
"request_count": {
|
|
"distinct_n": 8,
|
|
"max": 750.0,
|
|
"median": 285.5,
|
|
"min": 122.0,
|
|
"n": 36
|
|
},
|
|
"stream_bytes": {
|
|
"distinct_n": 6,
|
|
"max": 29431988.0,
|
|
"median": 25012856.5,
|
|
"min": 17449143.0,
|
|
"n": 6
|
|
},
|
|
"trace_duration_s": {
|
|
"distinct_n": 1,
|
|
"max": 60.0,
|
|
"median": 60.0,
|
|
"min": 60.0,
|
|
"n": 36
|
|
},
|
|
"trials": 36
|
|
},
|
|
"schema": "intervention-response-phase-aware-existing-v2",
|
|
"status": "COMPLETE",
|
|
"trajectory": {
|
|
"blocks": [
|
|
{
|
|
"end_s": 6.0,
|
|
"start_s": 0.0
|
|
},
|
|
{
|
|
"end_s": 12.0,
|
|
"start_s": 6.0
|
|
},
|
|
{
|
|
"end_s": 18.0,
|
|
"start_s": 12.0
|
|
}
|
|
],
|
|
"features": {
|
|
"batch_size.mean": {
|
|
"block_medians": [
|
|
6.414992146729492,
|
|
7.937718404518684,
|
|
7.938679599969923
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 44.717667092379735,
|
|
"median": 2.4047228794539084,
|
|
"min": 0.12661717921526972,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 44.717667092379735,
|
|
"median": 1.715194194802789,
|
|
"min": -6.818300947246247,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.7466021228318502
|
|
},
|
|
"batch_tokens.mean": {
|
|
"block_medians": [
|
|
182.10234474017744,
|
|
172.7298136645963,
|
|
202.28603044732077
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 908.0537249893571,
|
|
"median": 75.02808983931541,
|
|
"min": 7.692264317468613,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 908.0537249893571,
|
|
"median": 21.47667280168706,
|
|
"min": -316.4167321876521,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.7254256632404311
|
|
},
|
|
"decode_batch_size.mean": {
|
|
"block_medians": [
|
|
6.354735422986181,
|
|
7.873722548380083,
|
|
7.875311697892343
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 44.34555129842486,
|
|
"median": 2.4157231102242664,
|
|
"min": 0.15471898197242862,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 44.34555129842486,
|
|
"median": 1.7131997250467412,
|
|
"min": -6.730278183383879,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.7458594110784589
|
|
},
|
|
"graph_full_share": {
|
|
"block_medians": [
|
|
0.9440933791043857,
|
|
0.9375353988798221,
|
|
0.9361517143943923
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.20131971051511277,
|
|
"median": 0.027226282812290892,
|
|
"min": 0.00023443249531129862,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.08227563742558686,
|
|
"median": -0.0017697738257967388,
|
|
"min": -0.20131971051511277,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.6826036350375567
|
|
},
|
|
"graph_none_share": {
|
|
"block_medians": [
|
|
0.05080178541907753,
|
|
0.05845283871165627,
|
|
0.06118872118872119
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.19429544487015749,
|
|
"median": 0.01965769475771631,
|
|
"min": 0.0028422907600285444,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.19429544487015749,
|
|
"median": 0.0040749575702726006,
|
|
"min": -0.0846718334643753,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.65716602522967
|
|
},
|
|
"graph_padding_fraction": {
|
|
"block_medians": [
|
|
0.005497687293445758,
|
|
0.0075248839288535735,
|
|
0.006051201250183472
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.014587933688096199,
|
|
"median": 0.002744262881772697,
|
|
"min": 4.7975018144524526e-05,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.014587933688096199,
|
|
"median": -0.0005530753213689365,
|
|
"min": -0.008320005718079814,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.3098612628425477
|
|
},
|
|
"kv_usage_end_minus_start": {
|
|
"block_medians": [
|
|
0.023398338975546773,
|
|
-0.001777370408002521,
|
|
0.0038338466257342163
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.262813467878909,
|
|
"median": 0.017891096563429187,
|
|
"min": 0.0015150201237521532,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.07218503346987704,
|
|
"median": -0.01752170592552832,
|
|
"min": -0.262813467878909,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": -0.012340699297270389
|
|
},
|
|
"kv_usage_max": {
|
|
"block_medians": [
|
|
0.03336296364308622,
|
|
0.04022944403944284,
|
|
0.030052448221737305
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.29553401938255563,
|
|
"median": 0.011325832339947661,
|
|
"min": 0.0002397506593143639,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.29553401938255563,
|
|
"median": 0.00083657676867116,
|
|
"min": -0.09986012588670201,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.4134682913789121
|
|
},
|
|
"kv_usage_mean": {
|
|
"block_medians": [
|
|
0.016731908775651918,
|
|
0.02129250131291721,
|
|
0.01914443130373699
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.271100737119201,
|
|
"median": 0.009396606492664827,
|
|
"min": 0.0007061033361193698,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.271100737119201,
|
|
"median": 0.002361521286958815,
|
|
"min": -0.06598177225454588,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.10554346287202872
|
|
},
|
|
"preemptions": {
|
|
"block_medians": [
|
|
0.0,
|
|
0.0,
|
|
0.0
|
|
],
|
|
"changed_trials": 0,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"median": 0.0,
|
|
"min": 0.0,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 1,
|
|
"max": 0.0,
|
|
"median": 0.0,
|
|
"min": 0.0,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": null
|
|
},
|
|
"prefill_token_fraction": {
|
|
"block_medians": [
|
|
0.9617744444596128,
|
|
0.955926466887596,
|
|
0.9488462859261657
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.0574568246654229,
|
|
"median": 0.013436705979377572,
|
|
"min": 0.00024055263508171443,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 0.008129115874074722,
|
|
"median": -0.013436705979377572,
|
|
"min": -0.0574568246654229,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.23303776424770067
|
|
},
|
|
"queue_running_mean": {
|
|
"block_medians": [
|
|
6.85322045575,
|
|
8.0,
|
|
8.0
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 44.04634992016667,
|
|
"median": 2.4438409930833327,
|
|
"min": 0.5856487165000002,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 44.04634992016667,
|
|
"median": 1.8893361798333332,
|
|
"min": -4.201522480666667,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.8338308544566243
|
|
},
|
|
"queue_waiting_mean": {
|
|
"block_medians": [
|
|
0.0,
|
|
0.020209624833333332,
|
|
0.008993971666666666
|
|
],
|
|
"changed_trials": 23,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 24,
|
|
"max": 39.884282579499995,
|
|
"median": 0.008993971666666666,
|
|
"min": 0.0,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 24,
|
|
"max": 39.884282579499995,
|
|
"median": 0.00026988233333333,
|
|
"min": -0.7860144426666666,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.8749679224022264
|
|
},
|
|
"scheduler_steps_per_s": {
|
|
"block_medians": [
|
|
75.83333333333334,
|
|
78.16666666666667,
|
|
73.58333333333334
|
|
],
|
|
"changed_trials": 36,
|
|
"first_to_last_abs_delta": {
|
|
"distinct_n": 36,
|
|
"max": 76.5,
|
|
"median": 28.250000000000004,
|
|
"min": 2.1666666666666714,
|
|
"n": 36
|
|
},
|
|
"first_to_last_delta": {
|
|
"distinct_n": 36,
|
|
"max": 73.33333333333333,
|
|
"median": -3.1666666666666714,
|
|
"min": -76.5,
|
|
"n": 36
|
|
},
|
|
"first_to_last_pearson": 0.40161538799446195
|
|
}
|
|
},
|
|
"trial_count": 36
|
|
}
|
|
}
|