#!/usr/bin/env python3 """ eval_task1_claude.py — Claude-as-judge evaluator for Task 1 (Ablation Objective Identification). Protocol (v18 rubric, 2026-05-29): v18 changes vs v17 (recall side, eval-only — RL reward still on v17): * 1.0 adds SAME-COMPONENT-DIFFERENT-ANGLE rule — TM names the same paper artifact as the focus and RQ probes a different but legitimate angle → 1.0 (was 0). Constraint: component must be paper-specific. * 0.5 adds Case B (TM precise + RQ generic = 0.5, was 1.0 effectively). * 0.5 adds HARD PREREQ: at least one of TM/RQ must mention a paper-specific term the focus also references. Pure paraphrase → 0. * 0.0 adds rule (d): TM+RQ both generic-category = 0 (was sometimes 0.5). Goal: widen strong/weak spread by promoting genuine same-component matches (helped R1 / V3.2 framing diversity) while killing the "vague TM + vague RQ that happens to share a keyword with the focus" 0.5 rescue that inflates weak-model recall. v17-aligned protocol (kept for reference): This eval mirrors the live RL judge prompt shipped in reward_part/azure_task1_judge/task1_azure_reward.py (v15/v16/v17 share the same prompt). Differences vs the RL judge are intentional: * eval input has no , so that block is dropped * judge output is JSON (Claude returns JSON cleanly) instead of per-candidate XML * server-side enforce_bipartite is reproduced post-hoc in parse_match_response (score-desc + gt-asc → 1:1 dedup) * count_penalty parameters are kept independent of the RL reward (see below) Scoring rubric (identical wording to v17 judge): * 1.0 — DEDICATED ATOMIC MATCH (includes comparison atomization and joint composite TMs that name both sides) * 0.5 — NEAR-1.0 WITH MINOR PHRASING FLAW; vague TM rescued by precise RQ, OR one-side-of-true-joint rescued by RQ. UMBRELLA TMs cannot be rescued by RQ — those score 0. * 0.0 — NO MATCH (umbrella TM / parenthetical-only mention / no unconsumed bullet addresses the focus / vague catchall / different aspect) Default judge model is claude-sonnet-4-6. Usage: python eval_task1_claude.py --input python eval_task1_claude.py --input --workers 5 --limit 50 """ import os import sys import json import re import shutil import subprocess import argparse import tempfile import atexit import time import concurrent.futures from threading import Lock from typing import Optional, Tuple, List, Dict # Share count_penalty + utils with reward_part/azure_task1_judge so RL and # eval can never drift apart. _REWARD_PART_DIR = os.path.normpath( os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "reward_part")) if _REWARD_PART_DIR not in sys.path: sys.path.append(_REWARD_PART_DIR) from task1_candidate_utils import ( # noqa: E402 COUNT_PENALTY_BASE_RATE_DEFAULT, COUNT_PENALTY_CAP_DEFAULT, COUNT_PENALTY_HARD_CAP_DEFAULT, COUNT_PENALTY_STEP_DEFAULT, compute_count_penalty, compute_count_penalty_from_env, ) # ======================== Config ======================== CLAUDE_TIMEOUT = 300 WORKERS = 3 CLAUDE_MODEL: Optional[str] = None write_lock = Lock() # Throwaway cwd for `claude` subprocess. The CLI's background ai-title task # writes a ~130-byte stub to ~/.claude/projects// even with # --no-session-persistence (the flag blocks full transcripts but not the # title metadata stub). By running claude with cwd=this dir, every stub # lands in an isolated project directory we delete on exit. _SESSION_TMP_DIR = tempfile.mkdtemp(prefix="abforge_eval_t1_") def _mangle_project_path(p: str) -> str: # Claude Code maps cwd "/a/b c" -> "~/.claude/projects/-a-b-c" return "-" + p.lstrip("/").replace("/", "-").replace(" ", "-") def _cleanup_session_dir() -> None: proj_dir = os.path.expanduser( f"~/.claude/projects/{_mangle_project_path(_SESSION_TMP_DIR)}" ) shutil.rmtree(proj_dir, ignore_errors=True) shutil.rmtree(_SESSION_TMP_DIR, ignore_errors=True) atexit.register(_cleanup_session_dir) # ======================== Evaluation Prompt (v17-aligned) ======================== # # Direct adaptation of reward_part/azure_task1_judge/task1_azure_reward.py # create_evaluation_prompt (live v15/v16/v17 judge prompt). Three eval-side # adaptations: # 1. No — eval inputs do not carry it. # 2. JSON output (Claude handles JSON cleanly) instead of per-candidate XML. # 3. Bipartite 1:1 enforce is done post-hoc by parse_match_response. EVAL_PROMPT = """You are a rigorous scientific reviewer evaluating ablation objective predictions against GT focuses. {GT_OBJECTIVES} {PRED_OBJECTIVES} EVALUATION PROCEDURE: Process the reference focuses 1..N IN ORDER. For each focus: 1. Read the focus. 2. Look at the UNCONSUMED candidate bullets (bullets not yet cited as primary_match by an earlier focus). 3. For each candidate, judge whether its (Target Module + Research Question) addresses the same atomic ablation as the focus. 4. Assign the score 0 / 0.5 / 1.0 per the rubric below. Record the bullet's integer ID in primary_match (or 0 if no bullet qualifies). 5. A bullet picked as primary_match for the current focus is CONSUMED — do NOT cite it for any later focus. ATOMIC ABLATION means: a single specific paper-introduced method, component, design choice, or experimental contrast (e.g., "Beam search width", "Prototype Attentive Module", "BiasNorm vs LayerNorm", "L_adv + L_unsup joint loss"). NOT a combo of abstract categories. DUAL SIGNAL: Use BOTH the bullet's Target Module heading AND its Research Question text. TM tells you the bullet's main subject; RQ confirms (or disconfirms) it engages the focus's actual experimental concern. A bullet earns credit only when TM + RQ together point to the focus. SCORING (v18 — tightened 0.5 + same-component 1.0): 1.0 — ATOMIC MATCH ON SAME PAPER COMPONENT Award 1.0 in ANY of these cases (component identity matters more than verb identity): • DIRECT MATCH: TM clearly names the same atomic ablation as the focus AND RQ confirms the same experimental concern. Synonymous verbs are fine ("necessity" ≈ "contribution" ≈ "effect" ≈ "with-vs-without"). • SINGLE-COMPONENT ATOMIZATION (Ex: focus="Effect of beam width on translation" ←→ TM="Beam search width" → 1.0) • COMPARISON ATOMIZATION: focus="A vs B" is ONE experiment; a bullet naming A, naming B, or naming the contrast all score 1.0. (Ex: focus="LoRA vs Fully Fine-tuning" ←→ TM="LoRA fine-tuning" with RQ about replacing full FT → 1.0) • JOINT COMPOSITE: focus names X AND Y as a paper-joint ablation, TM names BOTH. (Ex: focus="L_adv and L_unsup loss design" ←→ TM="L_adv + L_unsup joint loss" → 1.0) • SAME COMPONENT, DIFFERENT VALID ANGLE (v18 — NEW): TM names the same paper-specific component/mechanism as the focus, AND the RQ probes a different but legitimate experimental angle on that same component (sensitivity ↔ robustness ↔ necessity ↔ magnitude). The bullet must be testing the same paper artifact, just from another direction. (Ex: focus="Effect of Reference Corpus Size on NPM" ←→ TM="Reference Corpus Invariance" with RQ asking if model adapts when corpus is swapped → 1.0, because both ablate the reference corpus.) CONSTRAINT: This rule applies ONLY when the component is paper-specific (named module / mechanism / design choice). It does NOT promote generic-category bullets — "Encoder Architecture" or "Loss Function" cannot use this rule. 0.5 — RESCUED MATCH (v18 — TIGHTENED, USE SPARINGLY) HARD PREREQUISITE: at least one of TM or RQ MUST mention a paper-specific term that the focus also references (a named module, mechanism, design choice, or technical detail visible in the focus or its description). If the bullet is generic enough to appear unchanged in any ML paper, score 0 — NOT 0.5. RQ keyword overlap with the focus alone is NOT sufficient — there must be paper-specific anchoring. Award 0.5 in the following cases (prereq must hold first): Case A — Vague TM rescued by precise RQ: TM is broader/vaguer than the focus, BUT the RQ contains a paper-specific term that locks the bullet to the focus's atomic concern. (Ex: focus="Effect of beam width" ←→ TM="Decoding parameter tuning" with RQ explicitly varying beam width on the paper's decoder → 0.5) Case B — Precise TM but generic RQ (v18 — NEW): TM exactly names the focus's paper-specific component, BUT the RQ is a generic template ("Is X critical for performance?" / "Does X improve results?") that adds no paper-specific detail beyond what the TM already says. The bullet shows it knows what to ablate but not how. (Ex: focus="MMD-based feature drift constraint" ←→ TM="MMD-based feature drift constraint" with RQ="Is the MMD constraint critical for performance?" → 0.5, because TM nails the component but RQ contributes no additional paper-specific anchor.) Case C — One side of TRUE joint/comparison: TM names ONE side of a TRUE joint ablation OR ONE side of a TRUE comparison, AND the RQ confirms the bullet IS engaging the focus's experimental point. (Ex: focus="Contribution of L_adv and L_unsup joint loss" ←→ TM="Adversarial loss" with RQ about removing GAN loss only → 0.5) Do NOT use 0.5 for "topically related" / "different aspect" / "higher abstraction" / "adjacent question" — those are 0. 0.0 — NO MATCH. Apply 0 if ANY hold: (a) The bullet's Target Module is a GENERIC UMBRELLA — combines 2+ abstract method categories rather than naming paper-specific components. UMBRELLA examples (always 0): "Training Protocol and Optimization Strategy" / "Backbone Architecture and Network Design" "Loss Function and Regularization" / "Ablation and Sensitivity Analysis" "Reward and Exploration Signal Design" / "Mechanistic Analysis" NOT-UMBRELLA (paper-named, may qualify): "MeanNet and BiasNet" / "L_adv and L_unsup loss" / "Planning Algorithm (MCTS and P-UCB)" TEST: remove "and"/"with". Are remaining noun phrases SPECIFIC paper names? → not umbrella. Are they generic categories appearing in any paper? → umbrella. (b) The focus appears only inside a parenthetical enumeration "(e.g., A, B, ...)" of a bullet whose TM is about a different topic. (c) No unconsumed bullet specifically addresses the focus (different aspect / different component / vague catchall listing many components). (d) (v18 — NEW) PURE PARAPHRASE: Both TM and RQ use only generic-category phrasings that could appear in many ML papers, even if topically aligned with the focus. There is no paper-specific anchor in either field — score 0, not 0.5. 1:1 ENFORCEMENT (post-hoc by evaluator): After your scoring, the evaluator script keeps each bullet's pairing only with the HIGHEST-scoring focus (tie-break: earlier focus wins). If you cite the same bullet for two focuses, the lower-scoring focus will be zeroed. Follow the procedure above strictly — pick the next-best unconsumed bullet, or primary_match=0. **Output Format (strictly this JSON, no markdown fences, no commentary):** {{ "matches": [ {{"gt_id": 1, "primary_match": 3, "score": 1.0, "reason": ""}} ], "unmatched_gt": [2] }}""" # ======================== IO ======================== def load_jsonl(filename): data = [] if os.path.exists(filename): with open(filename, 'r', encoding='utf-8') as f: for line in f: try: data.append(json.loads(line)) except Exception: pass return data def append_to_jsonl(filename, data): with write_lock: with open(filename, 'a', encoding='utf-8') as f: f.write(json.dumps(data, ensure_ascii=False) + "\n") f.flush() def get_title(item): return (item.get("meta", {}).get("title", "") or "").strip() def get_done_titles(*files): done = set() for f in files: for d in load_jsonl(f): t = (d.get("meta") or {}).get("title") if t: done.add(t.strip()) return done # ======================== Parsing ======================== # Same regex shape as the reward server (reward_part/.../task1_azure_reward.py) # so eval bullet count matches reward n_pairs telemetry. # Leading bullet char [-*] is optional to handle models (e.g. deepseek-v3.2) # that emit "**Target Module:** ..." without a dash prefix. _TARGET_BULLET_RE = re.compile( r'\n\s*(?:#{1,6}\s*)?(?:[-*]\s*)?(?:\*{1,2})?\s*Target Module\s*(?::\s*(?:\*{1,2})?|\*{1,2}\s*:)', re.IGNORECASE, ) _RQ_BULLET_RE = re.compile( r'\n\s*(?:#{1,6}\s*)?(?:[-*]\s*)?(?:\*{1,2})?\s*Research Question\s*(?::\s*(?:\*{1,2})?|\*{1,2}\s*:)', re.IGNORECASE, ) def parse_gt_objectives(gt_candidates: str) -> List[str]: """Extract each from the gt_Candidates XML.""" focuses = re.findall( r'(.*?)', gt_candidates, re.DOTALL, ) return [f.strip() for f in focuses if f.strip()] def parse_predicted_bullets(infer_response: str) -> List[Dict[str, str]]: """Parse infer response into ordered (target_module, research_question) bullets. Mirrors reward_part/.../task1_azure_reward.py:extract_numbered_bullets so the eval judge sees TM and RQ as separate signals (DUAL SIGNAL rubric requires this). RQ body is truncated to 400 chars to keep the judge prompt bounded — base/SFT/RL outputs all stay under 280 chars in practice (verified on bench200), so this rarely fires. """ if not infer_response: return [] # Pull ... if present, otherwise treat the entire text as the result. result_match = re.search(r'(.*?)(?:|$)', infer_response, re.DOTALL) result_text = result_match.group(1) if result_match else infer_response text = "\n" + result_text tm_matches = list(_TARGET_BULLET_RE.finditer(text)) rq_matches = list(_RQ_BULLET_RE.finditer(text)) bullets: List[Dict[str, str]] = [] for i, tm in enumerate(tm_matches): tm_start = tm.end() next_tm_start = tm_matches[i + 1].start() if i + 1 < len(tm_matches) else len(text) rq_in_range = next((rq for rq in rq_matches if tm_start < rq.start() < next_tm_start), None) if rq_in_range is None: continue tm_text = text[tm_start:rq_in_range.start()].strip().lstrip(":").strip().strip("*").strip() rq_text = text[rq_in_range.end():next_tm_start].strip().lstrip(":").strip().strip("*").strip() if len(rq_text) > 400: rq_text = rq_text[:400] + "..." bullets.append({ "idx": len(bullets) + 1, "target_module": tm_text[:300], "research_question": rq_text, }) return bullets def format_pred_bullets(bullets: List[Dict[str, str]]) -> str: if not bullets: return "(no parseable bullets in the generated result)" parts = [] for b in bullets: parts.append( f'\n' f'{b["target_module"]}\n' f'{b["research_question"]}\n' f'' ) return "\n".join(parts) def parse_match_response(response: str, n_gt: int, n_pred: int) -> Dict: """Parse judge JSON and apply bipartite 1:1 enforce (score-desc + gt-asc). Mirrors reward_part/.../task1_azure_reward.py:enforce_bipartite — when two focuses cite the same bullet, keep the higher-scoring focus; tie-break by earlier focus position. Replaces the older `seen_gt`/`seen_pred` dedup that was order-dependent on the judge's output order. """ json_match = re.search(r'\{.*\}', response, re.DOTALL) if not json_match: return {"matches": [], "parse_error": "no JSON found", "raw": response[-500:]} try: result = json.loads(json_match.group(0)) except json.JSONDecodeError as e: return {"matches": [], "parse_error": str(e), "raw": response[-500:]} raw_matches = result.get("matches") or result.get("scores") or [] # First pass: validate ranges, drop garbage, accept score==0 (kept for telemetry, # filtered later by score>0 when computing recall). candidates: List[Dict] = [] for m in raw_matches: gt_id = m.get("gt_id") # Accept either primary_match (v17) or pred_id (legacy) as the bullet pointer. pred_id = m.get("primary_match") if pred_id is None: pred_id = m.get("pred_id") score = m.get("score", 0) if not isinstance(gt_id, int) or gt_id < 1 or gt_id > n_gt: continue if not isinstance(score, (int, float)): continue # Normalize to {0, 0.5, 1.0}. score = float(score) score = min((0.0, 0.5, 1.0), key=lambda a: abs(a - score)) # primary_match == 0 (or out of range) means "no bullet for this GT". if not isinstance(pred_id, int) or pred_id < 1 or pred_id > n_pred: pred_id = 0 score = 0.0 candidates.append({ "gt_id": gt_id, "pred_id": pred_id, "score": score, "reason": m.get("reason", ""), }) # Second pass: bipartite enforce — keep highest score per pred_id, then # one match per gt_id. Sort by (score desc, gt_id asc); both kept-checks # operate on the same sort order so the score-winning focus survives even # if the judge listed it later. candidates.sort(key=lambda x: (-x["score"], x["gt_id"])) seen_gt = set() seen_pred = set() valid: List[Dict] = [] for c in candidates: if c["score"] <= 0 or c["pred_id"] == 0: continue if c["gt_id"] in seen_gt or c["pred_id"] in seen_pred: continue seen_gt.add(c["gt_id"]) seen_pred.add(c["pred_id"]) valid.append(c) # Re-sort matches by gt_id for stable downstream presentation. valid.sort(key=lambda x: x["gt_id"]) result["matches"] = valid return result def compute_match_rate(matches: List[Dict], n_gt: int) -> Optional[float]: if n_gt == 0: return None return sum(m.get("score", 1.0) for m in matches) / n_gt # Count penalty is the shared hard-cap function from reward_part. Defaults # (hard_cap=6, base=0.02, step=0.01, cap=0.5) are pulled from utils so RL # reward and eval can never drift. Override per-run with TASK1_COUNT_* # env vars. _eval_count_penalty = compute_count_penalty_from_env # ======================== Claude CLI ======================== def call_claude(prompt_text: str, label: str = "", timeout: int = CLAUDE_TIMEOUT) -> Optional[str]: tag = f"[{label}] " if label else "" cmd = [ "claude", "-p", "--output-format", "text", "--max-turns", "1", "--no-session-persistence", ] if CLAUDE_MODEL: cmd.extend(["--model", CLAUDE_MODEL]) try: t0 = time.time() proc = subprocess.run( cmd, input=prompt_text, capture_output=True, text=True, timeout=timeout, cwd=_SESSION_TMP_DIR, ) elapsed = time.time() - t0 if proc.returncode != 0: print(f" {tag}exit code {proc.returncode} ({elapsed:.1f}s)") if proc.stderr: print(f" stderr: {proc.stderr[:500]}") return None output = proc.stdout.strip() if not output: print(f" {tag}empty response ({elapsed:.1f}s)") return None print(f" {tag}ok ({elapsed:.1f}s, {len(output)} chars)") return output except subprocess.TimeoutExpired: print(f" {tag}timed out ({timeout}s)") return None except FileNotFoundError: print(f" {tag}'claude' not found in PATH") sys.exit(1) except Exception as e: print(f" {tag}error: {e}") return None # ======================== Per-item evaluation ======================== def evaluate_item(item: dict) -> Tuple[str, dict]: title = get_title(item) or "?" tag = title[:30] meta = item.get("meta", {}) try: gt_candidates = item.get("gt_Candidates", "") infer_response = item.get("infer_task1_response", "") if not gt_candidates or not infer_response: return ("FAIL", {"meta": meta, "reason": "missing gt_Candidates or infer_task1_response"}) gt_objectives = parse_gt_objectives(gt_candidates) if not gt_objectives: return ("FAIL", {"meta": meta, "reason": "no GT objectives parsed from gt_Candidates"}) pred_bullets = parse_predicted_bullets(infer_response) if not pred_bullets: return ("SUCCESS", { "meta": meta, "n_gt": len(gt_objectives), "n_pred": 0, "gt_objectives": gt_objectives, "pred_bullets": [], "matches": [], "unmatched_gt": list(range(1, len(gt_objectives) + 1)), "n_matched_full": 0, "n_matched_partial": 0, "n_matched": 0, "weighted_match_sum": 0.0, "match_rate": 0.0, "count_penalty": 0.0, "adjusted_score": 0.0, "raw_eval_response": "", "eval_note": "no predicted bullets parsed from infer_task1_response; scored as zero", }) gt_block = "\n".join( f'{obj}' for i, obj in enumerate(gt_objectives) ) pred_block = format_pred_bullets(pred_bullets) prompt = (EVAL_PROMPT .replace("{GT_OBJECTIVES}", gt_block) .replace("{PRED_OBJECTIVES}", pred_block)) resp = call_claude(prompt, label=f"EVALt1|{tag}") if not resp: return ("FAIL", {"meta": meta, "reason": "claude call failed"}) n_gt = len(gt_objectives) n_pred = len(pred_bullets) match_result = parse_match_response(resp, n_gt, n_pred) if "parse_error" in match_result: return ("FAIL", { "meta": meta, "reason": f"JSON parse error: {match_result['parse_error']}", "raw_response_tail": resp[-500:], }) matches = match_result.get("matches", []) n_matched_full = sum(1 for m in matches if m.get("score", 0) >= 0.99) n_matched_partial = sum(1 for m in matches if 0.3 < m.get("score", 0) < 0.99) weighted_match_sum = sum(m.get("score", 0) for m in matches) recall = weighted_match_sum / n_gt if n_gt else 0.0 count_penalty = _eval_count_penalty(n_pred) adjusted_score = max(0.0, recall - count_penalty) return ("SUCCESS", { "meta": meta, "n_gt": n_gt, "n_pred": n_pred, "gt_objectives": gt_objectives, "pred_bullets": pred_bullets, "matches": matches, "unmatched_gt": match_result.get("unmatched_gt", []), "n_matched_full": n_matched_full, "n_matched_partial": n_matched_partial, "n_matched": len(matches), "weighted_match_sum": weighted_match_sum, "match_rate": recall, "count_penalty": count_penalty, "adjusted_score": adjusted_score, "raw_eval_response": resp, }) except Exception as e: return ("FAIL", {"meta": meta, "reason": f"exception: {e}"}) # ======================== Runner ======================== def run_eval(input_file: str, output_file: str, fail_file: str, limit: int = 0): items = load_jsonl(input_file) if limit > 0: items = items[:limit] done = get_done_titles(output_file, fail_file) pending = [it for it in items if get_title(it) not in done] print(f"Task 1 Eval: total={len(items)}, done={len(done)}, " f"pending={len(pending)}, workers={WORKERS}") print(f"Model: {CLAUDE_MODEL or 'default'}") print(f"Input: {input_file}") print(f"Output: {output_file} / {fail_file}") if not pending: print("All done!") return success_count = fail_count = 0 with concurrent.futures.ThreadPoolExecutor(max_workers=WORKERS) as executor: future_map = { executor.submit(evaluate_item, item): item for item in pending } for future in concurrent.futures.as_completed(future_map): item = future_map[future] title = get_title(item) or "?" try: status, data = future.result() except Exception as e: status = "FAIL" data = {"meta": item.get("meta", {}), "reason": f"future exception: {e}"} if status == "SUCCESS": append_to_jsonl(output_file, data) success_count += 1 rate = data.get("match_rate") rate_str = f"{rate:.3f}" if rate is not None else "N/A" n_matched = data.get("n_matched", 0) n_gt = data.get("n_gt", 0) print(f" [{success_count + fail_count}/{len(pending)}] " f"{title[:50]} -> {rate_str} ({n_matched}/{n_gt})") else: append_to_jsonl(fail_file, data) fail_count += 1 reason = data.get("reason", "unknown") print(f" [{success_count + fail_count}/{len(pending)}] " f"{title[:50]} -> FAIL: {reason}") print(f"\n{'='*60}") print(f"Task 1 Eval done: {success_count} success, {fail_count} fail") # Final stats include FAILs as zero — silently dropping them inflates # over-generating models (v3 had 21/50 FAILs scoring 0.69 only because # the parse-failed cases were excluded from the average). ok_results = load_jsonl(output_file) fail_results = load_jsonl(fail_file) total_n = len(ok_results) + len(fail_results) if total_n == 0: return def avg_with_fails(key): vals = [r.get(key) or 0.0 for r in ok_results] vals.extend([0.0] * len(fail_results)) return sum(vals) / len(vals) if vals else 0.0 avg_recall = avg_with_fails("match_rate") avg_pen = avg_with_fails("count_penalty") avg_adj = avg_with_fails("adjusted_score") pred_counts = [r.get("n_pred", 0) for r in ok_results] gt_counts = [r.get("n_gt", 0) for r in ok_results] print(f"Total scored: {total_n} ({len(ok_results)} ok + {len(fail_results)} fail counted as 0)") print() hard_cap = int(os.environ.get("TASK1_COUNT_HARD_CAP", str(COUNT_PENALTY_HARD_CAP_DEFAULT))) base_rate = float(os.environ.get("TASK1_COUNT_BASE_RATE", str(COUNT_PENALTY_BASE_RATE_DEFAULT))) print(f" match_rate = {avg_recall:.4f} (raw recall, bipartite-enforced)") print(f" count_penalty = {avg_pen:.4f} (hard_cap={hard_cap}, base_rate={base_rate})") print(f" adjusted_score = {avg_adj:.4f} ← main metric (recall - count_penalty)") if pred_counts: print() print(f" n_pred mean={sum(pred_counts)/len(pred_counts):.2f} median={sorted(pred_counts)[len(pred_counts)//2]}") print(f" n_gt mean={sum(gt_counts)/len(gt_counts):.2f} median={sorted(gt_counts)[len(gt_counts)//2]}") n_full = sum(r.get("n_matched_full", 0) for r in ok_results) n_part = sum(r.get("n_matched_partial", 0) for r in ok_results) n_total_gt = sum(gt_counts) if n_total_gt: print(f" match mix 1.0={n_full} 0.5={n_part} 0={n_total_gt - n_full - n_part} " f"(of {n_total_gt} GT slots)") # ======================== Main ======================== if __name__ == "__main__": parser = argparse.ArgumentParser( description="Evaluate Task 1 inference results using Claude as judge") parser.add_argument("--input", required=True, help="Infer jsonl from run_inference_local.py --task 1") parser.add_argument("--output", default=None, help="Output jsonl (default: _eval_task1.jsonl)") parser.add_argument("--fail", default=None, help="Fail jsonl (default: _eval_task1_fail.jsonl)") parser.add_argument("--workers", type=int, default=3) parser.add_argument("--model", type=str, default="claude-sonnet-4-6", help="Claude model id (default: claude-sonnet-4-6)") parser.add_argument("--timeout", type=int, default=300) parser.add_argument("--limit", type=int, default=0, help="Only evaluate first N records (0 = all)") args = parser.parse_args() CLAUDE_MODEL = args.model CLAUDE_TIMEOUT = args.timeout WORKERS = args.workers if args.output is None: base = os.path.basename(args.input) stem = re.sub(r"_infer_task1\.jsonl$", "", base) if stem == base: stem = base[:-6] if base.endswith(".jsonl") else base out_dir = os.path.dirname(os.path.abspath(args.input)) args.output = os.path.join(out_dir, f"{stem}_eval_task1.jsonl") if args.fail is None: args.fail = args.output.replace(".jsonl", "_fail.jsonl") print(f"Config: model={CLAUDE_MODEL or 'default'}, " f"workers={WORKERS}, timeout={CLAUDE_TIMEOUT}s, limit={args.limit or 'all'}") run_eval(args.input, args.output, args.fail, limit=args.limit)