Datasets:
File size: 4,558 Bytes
46e27b5 | 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 | #!/usr/bin/env python3
"""Phase A — re-score the existing bakeoff under the DUAL Gate 2 constraint.
The stored benchmark JSONs already contain a complete threshold sweep with both
`over_deletion_rate` and `harmful_span_rate`, so re-selecting the operating
point is arithmetically identical to what the updated harness would produce,
without re-running inference.
Reports, per arm: which thresholds pass each gate, whether ANY threshold passes
both, and the resulting gap to target.
"""
import glob
import json
import os
import sys
sys.path.insert(0, "/opt/vox/sandbox/scripts")
from span_common import HARMFUL_SPAN_GATE, OVER_DELETION_GATE
OUT = "/opt/vox/sandbox/benchmarks/gate2_rescored.json"
def reselect(sweep):
for r in sweep:
r["passes_over_deletion_gate"] = r["over_deletion_rate"] < OVER_DELETION_GATE
r["passes_harmful_gate"] = r["harmful_span_rate"] < HARMFUL_SPAN_GATE
r["passes_gate"] = r["passes_over_deletion_gate"] and r["passes_harmful_gate"]
ok = [r for r in sweep if r["passes_gate"]]
if ok:
best = max(ok, key=lambda r: r["exact_match"])
best = dict(best)
best["no_threshold_satisfies_gate"] = False
else:
best = dict(min(sweep, key=lambda r: (r["harmful_span_rate"], r["over_deletion_rate"])))
best["no_threshold_satisfies_gate"] = True
return best, ok
def main():
report = {"gates": {"over_deletion_rate": OVER_DELETION_GATE,
"harmful_span_rate": HARMFUL_SPAN_GATE}, "arms": {}}
print(f"GATE 2 (dual): over_deletion < {OVER_DELETION_GATE}% AND "
f"harmful_span < {HARMFUL_SPAN_GATE}%\n")
hdr = (f"{'arm':22} {'thr':>5} {'EM':>7} {'P':>6} {'R':>6} {'F1':>6} "
f"{'overdel%':>9} {'harm%':>7} {'od':>3} {'harm':>5} both")
print(hdr)
print("-" * len(hdr))
for path in sorted(glob.glob("/opt/vox/sandbox/benchmarks/ft_*.json")):
tag = os.path.basename(path)[3:-5]
d = json.load(open(path))
best, ok = reselect(d["sweep"])
d["best"] = best
d["gates"] = report["gates"]
n_pass = len(ok)
# write the corrected selection back so the artifact matches the harness
json.dump(d, open(path, "w"), indent=2)
s = best
report["arms"][tag] = {
"params_m": d.get("params_m"), "cpu_p95_ms": d.get("cpu_latency", {}).get("p95_ms"),
"best": {k: s.get(k) for k in ("threshold", "exact_match", "span_word",
"over_deletion_rate", "harmful_span_rate",
"passes_gate", "no_threshold_satisfies_gate")},
"per_category": s.get("per_category"),
"thresholds_passing_both": n_pass,
"thresholds_passing_over_deletion_only":
sum(1 for r in d["sweep"] if r["passes_over_deletion_gate"]),
"thresholds_passing_harmful_only":
sum(1 for r in d["sweep"] if r["passes_harmful_gate"]),
"min_harmful_span_rate": min(r["harmful_span_rate"] for r in d["sweep"]),
"min_over_deletion_rate": min(r["over_deletion_rate"] for r in d["sweep"]),
}
print(f"{tag:22} {s['threshold']:>5} {s['exact_match']:>7} "
f"{s['span_word']['precision']:>6} {s['span_word']['recall']:>6} "
f"{s['span_word']['f1']:>6} {s['over_deletion_rate']:>9} "
f"{s['harmful_span_rate']:>7} "
f"{'Y' if s['passes_over_deletion_gate'] else 'n':>3} "
f"{'Y' if s['passes_harmful_gate'] else 'n':>5} "
f"{'YES' if s['passes_gate'] else 'NO':>4}")
print()
any_pass = [t for t, v in report["arms"].items() if v["best"]["passes_gate"]]
print(f"arms passing BOTH gates: {len(any_pass)}/{len(report['arms'])}"
f"{' -> ' + ', '.join(any_pass) if any_pass else ''}")
print("\nGap to target (best achievable per arm, i.e. most conservative threshold):")
print(f"{'arm':22} {'min harmful%':>13} {'target':>7} {'gap':>7} "
f"{'min overdel%':>13} {'target':>7}")
for tag, v in sorted(report["arms"].items(),
key=lambda kv: kv[1]["min_harmful_span_rate"]):
mh, mo = v["min_harmful_span_rate"], v["min_over_deletion_rate"]
print(f"{tag:22} {mh:>13} {HARMFUL_SPAN_GATE:>7} {round(mh - HARMFUL_SPAN_GATE, 3):>7} "
f"{mo:>13} {OVER_DELETION_GATE:>7}")
json.dump(report, open(OUT, "w"), indent=2)
print(f"\nwritten: {OUT}")
if __name__ == "__main__":
main() |