"""Eyeball what Model V actually outputs — to diagnose a weak ProcessBench score. python scripts/inspect_verifier.py --adapter runs/verifier_v --n 6 --only error Prints, per example: gold first-error index, V's parsed prediction, and the raw critique. Reveals whether V is collapsing to -1, mislocating, off-by-one, or rambling/truncating. """ import argparse import sys from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "src")) from mathcompose.common.chat import verifier_messages # noqa: E402 from mathcompose.data.schema import parse_verifier_output # noqa: E402 def main() -> int: ap = argparse.ArgumentParser() ap.add_argument("--base-id", default="Qwen/Qwen2.5-Math-1.5B-Instruct") ap.add_argument("--adapter", default=None) ap.add_argument("--split", default="gsm8k") ap.add_argument("--n", type=int, default=6) ap.add_argument("--only", choices=["error", "correct", "both"], default="error") ap.add_argument("--max-new-tokens", type=int, default=768) args = ap.parse_args() from datasets import load_dataset from mathcompose.eval.runners import HFRunner ds = load_dataset("Qwen/ProcessBench", split=args.split).shuffle(seed=0) runner = HFRunner(args.base_id, adapter_path=args.adapter) shown = 0 for row in ds: gold = int(row["label"]) if args.only == "error" and gold == -1: continue if args.only == "correct" and gold != -1: continue out = runner.chat(verifier_messages(row["problem"], list(row["steps"])), n=1, temperature=0.0, max_new_tokens=args.max_new_tokens)[0] pred = parse_verifier_output(out).first_error_index flag = "HIT" if pred == gold else "miss" print(f"\n{'=' * 70}\n[{flag}] gold={gold} pred={pred} (#steps={len(row['steps'])})") print("-" * 70) print(out.strip()[:1200]) shown += 1 if shown >= args.n: break return 0 if __name__ == "__main__": raise SystemExit(main())