File size: 8,582 Bytes
8b97eb8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 | #!/usr/bin/env python3
"""Rescore an existing baseline_agent batch run with the current numeric scorer.
The script reads a run's summary.tsv, scores every listed submission with
`harness/evaluate_numeric.py`, and writes:
- summary_rescored.tsv
- summary_rescored.json
- numeric_rescored/<model>/<type>/<task>.json
Completion means a qualified submission: the solver produced a submission that
compiled, passed the contract, executed, and received numeric_status == "ok".
"""
from __future__ import annotations
import argparse
import csv
import json
import math
import subprocess
import sys
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any
def _repo_root() -> Path:
return Path(__file__).resolve().parent.parent
def _as_float(value: Any) -> float | None:
if isinstance(value, bool):
return None
if isinstance(value, (int, float)):
out = float(value)
return out if math.isfinite(out) else None
try:
out = float(str(value))
except Exception:
return None
return out if math.isfinite(out) else None
def _score_zero(submission: Path, task_type: str, status: str, error: str = "") -> dict[str, Any]:
return {
"submission": submission.name,
"status": status,
"contract_ok": False,
"numeric_score": 0.0,
"raw_numeric_score": None,
"numeric_score_std": 0.0,
"numeric_score_per_seed": [0.0] if task_type == "typeI" else [0.0, 0.0, 0.0],
"raw_numeric_score_per_seed": [None] if task_type == "typeI" else [None, None, None],
"raw_metric": None,
"error": error,
}
def _score_submission(repo: Path, task_dir: Path, submission: Path, task_type: str) -> dict[str, Any]:
if not submission.exists():
return _score_zero(submission, task_type, "missing_submission", f"{submission} missing")
cmd = [
sys.executable,
str(repo / "harness" / "evaluate_numeric.py"),
"score",
str(task_dir),
str(submission),
]
proc = subprocess.run(cmd, cwd=repo, text=True, capture_output=True, check=False)
if proc.returncode != 0:
return _score_zero(
submission,
task_type,
"scorer_error",
proc.stderr.strip() or proc.stdout.strip(),
)
try:
result = json.loads(proc.stdout)
except Exception as exc:
return _score_zero(
submission,
task_type,
"invalid_scorer_json",
f"{type(exc).__name__}: {exc}; stdout={proc.stdout[:500]!r}",
)
if _as_float(result.get("numeric_score")) is None:
result["numeric_score"] = 0.0
result.setdefault("status", "numeric_null")
return result
def _summary_stats(rows: list[dict[str, Any]]) -> dict[str, Any]:
scores = [_as_float(row.get("numeric_score")) or 0.0 for row in rows]
submitted = sum(str(row.get("submission_exists", "")).lower() == "true" for row in rows)
qualified = sum(str(row.get("qualified_submission", "")).lower() == "true" for row in rows)
numeric_status = Counter(row.get("numeric_status", "") for row in rows)
agent_status = Counter(row.get("status", "") for row in rows)
return {
"n": len(rows),
"submitted": submitted,
"submission_rate": submitted / len(rows) if rows else None,
"qualified_submissions": qualified,
"completion_rate": qualified / len(rows) if rows else None,
"invalid_submissions": len(rows) - qualified,
"mean_score": sum(scores) / len(scores) if scores else None,
"zero_numeric": sum(score == 0.0 for score in scores),
"numeric_status": dict(sorted(numeric_status.items())),
"agent_status": dict(sorted(agent_status.items())),
}
def _group_stats(rows: list[dict[str, Any]], key: str) -> dict[str, Any]:
grouped: dict[str, list[dict[str, Any]]] = defaultdict(list)
for row in rows:
grouped[row.get(key, "")].append(row)
return {name: _summary_stats(items) for name, items in sorted(grouped.items())}
def rescore_batch(args: argparse.Namespace) -> dict[str, Any]:
repo = args.repo_root.resolve()
run_dir = args.run_dir.resolve()
summary_path = run_dir / args.summary
if not summary_path.exists():
raise SystemExit(f"{summary_path} missing")
rows = list(csv.DictReader(summary_path.open(), delimiter="\t"))
if not rows:
raise SystemExit(f"{summary_path} has no rows")
model = args.model or rows[0].get("model")
if not model:
raise SystemExit("--model is required when summary.tsv has no model column")
out_root = run_dir / args.numeric_out / model
new_rows: list[dict[str, Any]] = []
for idx, row in enumerate(rows, start=1):
task_type = row["type"]
task = row["task"]
task_dir = repo / "tasks" / task_type / task
submission = run_dir / "submissions" / model / task_type / f"{task}.py"
out_dir = out_root / task_type
out_dir.mkdir(parents=True, exist_ok=True)
result = _score_submission(repo, task_dir, submission, task_type)
(out_dir / f"{task}.json").write_text(
json.dumps(result, indent=2, sort_keys=True) + "\n",
encoding="utf-8",
)
score = _as_float(result.get("numeric_score")) or 0.0
raw_score = result.get("raw_numeric_score")
numeric_status = str(result.get("status") or ("ok" if result.get("contract_ok") else "unknown"))
qualified = numeric_status == "ok"
new_row = dict(row)
new_row["numeric_score"] = str(score)
new_row["numeric_status"] = numeric_status
new_row["raw_numeric_score"] = "" if raw_score is None else str(raw_score)
new_row["numeric_error"] = str(result.get("error") or result.get("note") or "")
new_row["submission_exists"] = str(submission.exists()).lower()
new_row["qualified_submission"] = str(qualified).lower()
new_rows.append(new_row)
if args.progress_every > 0 and idx % args.progress_every == 0:
print(f"{run_dir.name} {idx} / {len(rows)}", flush=True)
out_summary = run_dir / args.output_summary
fieldnames = list(rows[0].keys())
for extra in [
"numeric_status",
"raw_numeric_score",
"numeric_error",
"submission_exists",
"qualified_submission",
]:
if extra not in fieldnames:
fieldnames.append(extra)
with out_summary.open("w", newline="", encoding="utf-8") as fh:
writer = csv.DictWriter(fh, fieldnames=fieldnames, delimiter="\t")
writer.writeheader()
writer.writerows(new_rows)
stats = {
"run": str(run_dir),
"model": model,
"summary_source": str(summary_path),
"summary_rescored": str(out_summary),
"numeric_rescored_dir": str(out_root),
"overall": _summary_stats(new_rows),
"by_type": _group_stats(new_rows, "type"),
}
out_json = run_dir / args.output_json
out_json.write_text(json.dumps(stats, indent=2, sort_keys=True) + "\n", encoding="utf-8")
if args.replace_summary:
backup = run_dir / args.backup_summary
if not backup.exists():
backup.write_bytes(summary_path.read_bytes())
summary_path.write_bytes(out_summary.read_bytes())
stats["summary_replaced"] = str(summary_path)
stats["summary_backup"] = str(backup)
out_json.write_text(json.dumps(stats, indent=2, sort_keys=True) + "\n", encoding="utf-8")
print(json.dumps(stats, indent=2, sort_keys=True))
return stats
def main() -> int:
parser = argparse.ArgumentParser(description="Rescore a baseline batch run with current numeric scoring.")
parser.add_argument("run_dir", type=Path)
parser.add_argument("--model", default=None)
parser.add_argument("--repo-root", type=Path, default=_repo_root())
parser.add_argument("--summary", default="summary.tsv")
parser.add_argument("--output-summary", default="summary_rescored.tsv")
parser.add_argument("--output-json", default="summary_rescored.json")
parser.add_argument("--numeric-out", default="numeric_rescored")
parser.add_argument("--replace-summary", action="store_true")
parser.add_argument("--backup-summary", default="summary_pre_numeric_zero.tsv")
parser.add_argument("--progress-every", type=int, default=25)
args = parser.parse_args()
rescore_batch(args)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|