Evaluate Qwen30 prefill simulator fidelity

2026-07-17 03:11:45 +08:00
parent 97c2f34700
commit 3a59d5df96
20 changed files with 8280 additions and 28 deletions
--- a/runs/frontier-phase-factorial-v0/test_phase_factorial.py
+++ b/runs/frontier-phase-factorial-v0/test_phase_factorial.py
@@ -37,3 +37,9 @@ def test_grid_and_trace(tmp_path: Path) -> None:
    assert len(lines) == 4
    assert lines[1].split(",")[:3] == ["0.000000000000", "2048", "1"]
    assert lines[3].split(",")[:3] == ["0.500000000000", "2048", "1"]
+
+
+def test_kendall_tau_b() -> None:
+    analysis = load("analyze_qwen30_prefill_fidelity.py")
+    assert analysis.kendall_tau_b([1, 2, 3], [1, 2, 3])["kendall_tau_b"] == 1
+    assert analysis.kendall_tau_b([1, 2, 3], [3, 2, 1])["kendall_tau_b"] == -1