NegBioDB / scripts /run_l3_judge_validation.py
jang1563's picture
NegBioDB final: 4 domains, fully audited
6d1bbc7
#!/usr/bin/env python3
"""Run second-judge validation for L3 reasoning evaluation.
Selects 3 representative L3 runs (best/median/worst), re-judges them
with a second judge model (GPT-4o-mini), and saves scores for
inter-rater agreement analysis.
Avoids circularity: GPT-4o-mini predictions are excluded from the sample
since GPT-4o-mini is used as the second judge.
Usage:
python scripts/run_l3_judge_validation.py \
--judge-provider openai --judge-model gpt-4o-mini
Output:
results/llm/judge_validation/
gpt4o_mini_scores.jsonl — per-item second-judge scores
gpt4o_mini_meta.json — judge model info, timestamp
"""
import argparse
import json
import time
from datetime import datetime, timezone
from pathlib import Path
from negbiodb.llm_client import LLMClient
from negbiodb.llm_eval import L3_JUDGE_PROMPT, parse_l3_judge_scores
PROJECT_ROOT = Path(__file__).resolve().parent.parent
RESULTS_DIR = PROJECT_ROOT / "results" / "llm"
DATA_DIR = PROJECT_ROOT / "exports" / "llm_benchmarks"
L3_GOLD_FILE = DATA_DIR / "l3_reasoning_pilot.jsonl"
# 3 representative runs (avoiding GPT-4o-mini predictions to prevent circularity)
REPRESENTATIVE_RUNS = [
"l3_gemini-2-5-flash_3-shot_fs0", # best overall (~4.64)
"l3_qwen2-5-32b-instruct-awq_3-shot_fs0", # median overall (~3.68)
"l3_mistral-7b-instruct-v0-3_3-shot_fs0", # worst overall (~3.18)
]
def load_gold_records() -> dict[str, dict]:
"""Load L3 gold records indexed by question_id."""
records = {}
with open(L3_GOLD_FILE) as f:
for line in f:
rec = json.loads(line)
records[rec["question_id"]] = rec
return records
def load_predictions(run_dir: Path) -> list[dict]:
"""Load predictions from JSONL."""
preds = []
pred_path = run_dir / "predictions.jsonl"
with open(pred_path) as f:
for line in f:
preds.append(json.loads(line))
return preds
def main():
parser = argparse.ArgumentParser(description="Run second-judge L3 validation")
parser.add_argument(
"--judge-provider", default="openai",
choices=["openai", "gemini", "vllm", "anthropic"],
)
parser.add_argument("--judge-model", default="gpt-4o-mini")
parser.add_argument("--api-base", default=None)
parser.add_argument("--api-key", default=None)
parser.add_argument("--temperature", type=float, default=0.0)
parser.add_argument("--max-tokens", type=int, default=512)
parser.add_argument(
"--results-dir", type=Path, default=RESULTS_DIR,
)
args = parser.parse_args()
# Output directory
out_dir = args.results_dir / "judge_validation"
out_dir.mkdir(parents=True, exist_ok=True)
# Load gold records
print("Loading L3 gold records...")
gold_by_id = load_gold_records()
print(f" Loaded {len(gold_by_id)} gold records")
# Initialize second judge
judge_name = args.judge_model.replace("/", "-")
print(f"\nInitializing second judge: {args.judge_model} ({args.judge_provider})")
client = LLMClient(
provider=args.judge_provider,
model=args.judge_model,
api_base=args.api_base,
api_key=args.api_key,
temperature=args.temperature,
max_tokens=args.max_tokens,
)
# Resume support
scores_path = out_dir / f"{judge_name}_scores.jsonl"
completed = {}
if scores_path.exists():
with open(scores_path) as f:
for line in f:
rec = json.loads(line)
key = (rec["source_run"], rec["question_id"])
if rec.get("scores") is not None:
completed[key] = rec
print(f" Resume: {len(completed)} already judged")
# Collect all items to judge
items = []
for run_name in REPRESENTATIVE_RUNS:
run_dir = args.results_dir / run_name
if not run_dir.exists():
print(f" WARNING: Run not found: {run_name}, skipping")
continue
predictions = load_predictions(run_dir)
for pred_rec in predictions:
qid = pred_rec["question_id"]
if (run_name, qid) not in completed:
items.append((run_name, qid, pred_rec["prediction"]))
print(f"\n Total items to judge: {len(items)} "
f"(from {len(REPRESENTATIVE_RUNS)} runs)")
# Judge
start_time = time.time()
with open(scores_path, "a") as f:
for i, (run_name, qid, prediction) in enumerate(items):
gold = gold_by_id.get(qid)
if gold is None:
print(f" Warning: no gold for {qid}, skipping")
continue
prompt = L3_JUDGE_PROMPT.format(
compound_name=gold.get("compound_name", "Unknown"),
target_gene=gold.get("target_gene") or "Unknown",
target_uniprot=gold.get("target_uniprot", "Unknown"),
response=prediction,
)
try:
judge_response = client.generate(prompt)
except Exception as e:
print(f" Error judging {run_name}/{qid}: {e}")
judge_response = f"ERROR: {e}"
scores = parse_l3_judge_scores(judge_response)
score_rec = {
"source_run": run_name,
"question_id": qid,
"judge_response": judge_response,
"scores": scores,
}
f.write(json.dumps(score_rec, ensure_ascii=False) + "\n")
f.flush()
completed[(run_name, qid)] = score_rec
if (i + 1) % 10 == 0:
elapsed = time.time() - start_time
rate = (i + 1) / elapsed * 60
print(f" Progress: {i + 1}/{len(items)} ({rate:.1f}/min)")
elapsed = time.time() - start_time
n_valid = sum(1 for v in completed.values() if v.get("scores") is not None)
print(f"\nJudging complete: {elapsed:.0f}s, {n_valid}/{len(completed)} valid")
# Save metadata
meta = {
"judge_model": args.judge_model,
"judge_provider": args.judge_provider,
"representative_runs": REPRESENTATIVE_RUNS,
"n_total_items": len(completed),
"n_valid_scores": n_valid,
"elapsed_seconds": elapsed,
"timestamp": datetime.now(timezone.utc).isoformat(),
}
meta_path = out_dir / f"{judge_name}_meta.json"
with open(meta_path, "w") as f:
json.dump(meta, f, indent=2)
print(f"Metadata: {meta_path}")
if __name__ == "__main__":
main()