File size: 6,051 Bytes
1ab87df 4e1a75b 1ab87df 4e1a75b 1ab87df 4e1a75b 1ab87df 4e1a75b 1ab87df 4e1a75b 1ab87df | 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 | """Replay generated examples against the current CORP-ENV implementation.
The script keeps only trajectories that parse as `CorpAction`, run through
`CorpEnvironment`, terminate with `finalize`, and meet task-specific thresholds.
Example:
uv run python scripts/verify_examples.py \
--input data/raw/e1_m1_examples.jsonl \
--clean data/processed/e1_m1_clean.jsonl \
--rejected data/processed/e1_m1_rejected.jsonl \
--summary results/example_verification_summary.json
"""
from __future__ import annotations
import argparse
import json
import sys
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any, Dict, List
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from scripts._trajectory_utils import ( # noqa: E402
extract_actions,
read_jsonl,
replay_actions,
write_jsonl,
)
def verify_one(
example: Dict[str, Any],
strict_thresholds: bool,
strict_json: bool,
require_stepwise_deliberation: bool,
) -> Dict[str, Any]:
example_id = str(example.get("example_id") or example.get("id") or "unknown")
task_id = str(example.get("task_id") or example.get("task") or "")
if example.get("_load_error"):
return {
"example_id": example_id,
"task_id": task_id,
"status": "rejected",
"reject_reason": example["_load_error"],
"actions": [],
}
if not task_id:
return {
"example_id": example_id,
"task_id": task_id,
"status": "rejected",
"reject_reason": "missing_task_id",
"actions": [],
}
try:
actions = extract_actions(example)
except Exception as exc:
return {
"example_id": example_id,
"task_id": task_id,
"status": "rejected",
"reject_reason": f"action_extraction_failed: {exc}",
"actions": [],
}
try:
result = replay_actions(
example_id=example_id,
task_id=task_id,
actions=actions,
strict_thresholds=strict_thresholds,
strict_json=strict_json,
require_stepwise_deliberation=require_stepwise_deliberation,
)
except Exception as exc:
return {
"example_id": example_id,
"task_id": task_id,
"status": "rejected",
"reject_reason": f"replay_failed: {exc}",
"actions": actions,
}
return result.to_record()
def summarize(records: List[Dict[str, Any]]) -> Dict[str, Any]:
by_status = Counter(r["status"] for r in records)
by_reason = Counter(r.get("reject_reason") or "clean" for r in records)
by_task: Dict[str, Counter] = defaultdict(Counter)
for row in records:
by_task[row.get("task_id", "unknown")][row["status"]] += 1
clean = [r for r in records if r["status"] == "clean"]
return {
"total": len(records),
"by_status": dict(by_status),
"by_reject_reason": dict(by_reason),
"by_task": {task: dict(counts) for task, counts in by_task.items()},
"clean_avg_terminal_reward": (
round(sum(float(r.get("terminal_reward", 0.0)) for r in clean) / len(clean), 6)
if clean
else 0.0
),
"clean_avg_verifier_pass_rate": (
round(sum(float(r.get("verifier_pass_rate", 0.0)) for r in clean) / len(clean), 6)
if clean
else 0.0
),
"clean_avg_reasoning_steps": (
round(sum(float(r.get("reasoning_steps", 0.0)) for r in clean) / len(clean), 6)
if clean
else 0.0
),
"clean_phase_progression_ok_rate": (
round(sum(1.0 for r in clean if r.get("phase_progression_ok")) / len(clean), 6)
if clean
else 0.0
),
}
def main() -> None:
parser = argparse.ArgumentParser(description="Verify generated CORP-ENV examples.")
parser.add_argument("--input", required=True, help="Raw or normalized examples JSONL.")
parser.add_argument("--clean", default="data/processed/e1_m1_clean.jsonl")
parser.add_argument("--rejected", default="data/processed/e1_m1_rejected.jsonl")
parser.add_argument("--all-records", default="results/example_verification_all.jsonl")
parser.add_argument("--summary", default="results/example_verification_summary.json")
parser.add_argument(
"--lenient",
action="store_true",
help="Only require replay validity; do not apply task reward/pass thresholds.",
)
parser.add_argument(
"--strict-json",
action="store_true",
help="Require strict single-object JSON action format with canonical action keys.",
)
parser.add_argument(
"--require-stepwise-deliberation",
action="store_true",
help="Enforce task-specific step-wise deliberation checks (reasoning/conflict/resolution/phases).",
)
args = parser.parse_args()
records = [
verify_one(
example,
strict_thresholds=not args.lenient,
strict_json=args.strict_json,
require_stepwise_deliberation=args.require_stepwise_deliberation,
)
for example in read_jsonl(Path(args.input))
]
clean = [r for r in records if r["status"] == "clean"]
rejected = [r for r in records if r["status"] != "clean"]
write_jsonl(Path(args.clean), clean)
write_jsonl(Path(args.rejected), rejected)
write_jsonl(Path(args.all_records), records)
summary = summarize(records)
summary_path = Path(args.summary)
summary_path.parent.mkdir(parents=True, exist_ok=True)
summary_path.write_text(json.dumps(summary, indent=2, ensure_ascii=False), encoding="utf-8")
print(json.dumps(summary, indent=2, ensure_ascii=False))
print(f"\nClean examples: {args.clean}")
print(f"Rejected examples: {args.rejected}")
if __name__ == "__main__":
main()
|