File size: 4,866 Bytes
a37a9bc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 | #!/usr/bin/env python3
"""Unit tests for the RE-1 log parsers and the two-run comparison."""
from __future__ import annotations
import copy
import json
import sys
import unittest
from pathlib import Path
HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE.parent / "harness"))
import analyze # noqa: E402
import run_pipeline # noqa: E402
FIXTURES = HERE.parent / "fixtures"
class ParseTests(unittest.TestCase):
def test_parse_ppl(self) -> None:
text = (FIXTURES / "ppl-sample.log").read_text()
parsed = run_pipeline.parse_ppl(text)
self.assertEqual(parsed["ppl"], 15.2105)
self.assertEqual(parsed["ppl_err"], 0.98064)
self.assertIsNotNone(parsed["load_time_ms"])
def test_parse_ppl_missing_fails_closed(self) -> None:
with self.assertRaises(ValueError):
run_pipeline.parse_ppl("no numbers here\n")
def test_parse_hellaswag_takes_last_row(self) -> None:
text = (FIXTURES / "hellaswag-sample.log").read_text()
parsed = run_pipeline.parse_hellaswag(text)
self.assertEqual(parsed["hellaswag_tasks_scored"], 8)
self.assertEqual(parsed["hellaswag_pct"], 37.5)
self.assertAlmostEqual(parsed["hellaswag_lo"], 13.6844)
self.assertAlmostEqual(parsed["hellaswag_hi"], 69.4258)
def test_parse_hellaswag_missing_fails_closed(self) -> None:
with self.assertRaises(ValueError):
run_pipeline.parse_hellaswag("no table\n")
def test_parse_bench_combined_log(self) -> None:
text = (FIXTURES / "bench-sample.log").read_text()
parsed = run_pipeline.parse_bench(text)
self.assertAlmostEqual(parsed["pp512_ts"], 102.089135)
self.assertAlmostEqual(parsed["tg128_ts"], 11.720742)
def fake_run(run_id: str, scale: float = 1.0) -> dict:
"""Minimal run dict with the shape flatten() consumes."""
def arm(ppl: float, hs: float, rss: int, load: float, pp: float, tg: float) -> dict:
return {
"ppl": ppl,
"ppl_err": 0.01,
"hellaswag_pct": hs,
"hellaswag_lo": hs - 5.0,
"hellaswag_hi": hs + 5.0,
"hellaswag_tasks_scored": 200,
"ppl_peak_rss_kib": rss,
"load_time_ms": load,
"pp512_ts": pp,
"tg128_ts": tg,
}
models = {}
for key, ppl, hs, rss, load, pp, tg in (
("model-a", 15.0, 30.0, 1_500_000, 500.0, 100.0, 10.0),
("model-b", 25.0, 40.0, 900_000, 300.0, 200.0, 20.0),
):
models[key] = {
"model": key,
"f16": {"bytes": 1_000_000_000},
"q4_k_m": {"bytes": 400_000_000},
"arms": {
"f16": arm(ppl, hs, rss, load, pp, tg),
"q4_k_m": arm(ppl * scale, hs + 1.0, int(rss * 0.9), load * 0.9, pp * 1.1, tg * 1.2),
},
"bench": {
"pp512_ts_f16": pp,
"pp512_ts_q4": pp * 1.1,
"tg128_ts_f16": tg,
"tg128_ts_q4": tg * 1.2,
"pp_ratio_median": 1.1,
"tg_ratio_median": 1.2,
},
}
return {"run_id": run_id, "started_at": "2026-09-17T00:00:00Z", "models": models}
class AnalyzeTests(unittest.TestCase):
def test_identical_runs_pass(self) -> None:
run1 = fake_run("r1")
result = analyze.compare(run1, copy.deepcopy(run1))
self.assertEqual(result["decision"], "capability_proven")
self.assertEqual(result["failing"], [])
self.assertEqual(result["missing"], [])
def test_delta_tolerance_is_scaled_by_reference_arm(self) -> None:
run1 = fake_run("r1")
run2 = fake_run("r2", scale=1.03)
result = analyze.compare(run1, run2)
# ppl_delta for model-a is 0.45 in both runs (15.0 * 1.03 - 15.0 = 0.45);
# its denominator is the f16 baseline 15.0, so the 3% arm shift moves the
# delta by exactly 0.45, which is 3% of baseline -> must fail.
self.assertEqual(result["quantities"]["model-a.ppl_delta"]["denominator"], 15.0)
self.assertFalse(result["quantities"]["model-a.ppl_delta"]["pass"])
self.assertEqual(result["decision"], "capability_gap")
def test_missing_quantity_fails_closed(self) -> None:
run1 = fake_run("r1")
run2 = fake_run("r2")
del run2["models"]["model-b"]
result = analyze.compare(run1, run2)
self.assertEqual(result["decision"], "capability_gap")
self.assertTrue(any(k.startswith("model-b.") for k in result["missing"]))
def test_markdown_renders(self) -> None:
run1 = fake_run("r1")
md = analyze.to_markdown(analyze.compare(run1, copy.deepcopy(run1)))
self.assertIn("capability_proven", md)
self.assertIn("model-a.ppl_delta", md)
if __name__ == "__main__":
unittest.main()
|