"""Evaluator self-test: a gold-derived 'perfect' answer must score Q=10 and a junk answer ~0 for every evaluator type.""" import json import sys from collections import deque import pandas as pd sys.path.insert(0, "src"); sys.path.insert(0, ".") from llmroute.seqroute.evaluate import evaluate_row df = pd.read_parquet("/mnt/localssd/seqroute1m/queries/queries_20k.parquet") def bw_plan(p): mv = [] for s in p["init"]: for b in reversed(s): mv.append(f"move {b} to table") for s in p["goal"]: for i in range(1, len(s)): mv.append(f"move {s[i]} to {s[i - 1]}") return json.dumps(mv) def hanoi_plan(p): pegs = list(p["init"]) start = tuple(tuple(p["init"][x]) for x in pegs) goal = tuple(tuple(p["goal"][x]) for x in pegs) prev, q = {start: None}, deque([start]) while q: s = q.popleft() if s == goal: break for i in range(3): if not s[i]: continue d = s[i][-1] for j in range(3): if i != j and (not s[j] or s[j][-1] > d): t = list(map(list, s)); t[i].pop(); t[j].append(d) t = tuple(map(tuple, t)) if t not in prev: prev[t] = (s, [d, pegs[i], pegs[j]]); q.append(t) mv, s = [], goal while prev.get(s): s, m = prev[s]; mv.append(m) return json.dumps(mv[::-1]) def perfect(r): ev, pl = r.evaluator_type, json.loads(r.payload) if ev == "math_exact": return f"Final answer: \\boxed{{{pl['gold']}}}" if ev == "mc_exact": return f"Reasoning...\nFinal answer: {pl['gold']}" if ev == "unit_tests": if pl["kind"] == "humaneval": ref = r.reference return f"```python\n{pl['func_prompt'] if not ref.lstrip().startswith(('def', 'from', 'import')) else ''}{ref}\n```" return f"```python\n{r.reference}\n```" if r.reference else None if ev == "extraction_f1": return r.reference if ev == "exact_str": return f"Final answer: ({pl['gold']})" if pl["kind"] == "letter" else f"Final answer: **{pl['gold']}**" if ev in ("label_exact",): return pl["gold"] if ev == "json_slots": return "```json\n" + r.reference + "\n```" if ev == "qa_f1": return r.reference if ev == "plan_blocksworld": return "Final answer: " + bw_plan(pl) if ev == "plan_hanoi": return "Final answer: " + hanoi_plan(pl) if pl["n_disks"] <= 5 else None if ev == "sudoku": return "Final answer:\n" + "\n".join("".join(map(str, row)) for row in pl["solution"]) return None bad = 0 for ev, g in df.groupby("evaluator_type"): if ev in ("judge_summ", "comet", "simplification"): continue res = [] for r in g.sample(min(40, len(g)), random_state=0).itertuples(): pa = perfect(r) pl = json.loads(r.payload) wrong = evaluate_row(ev, r.prompt, pl, "I am not sure. Final answer: zzz qq", None, "stop") if pa is None: res.append((None, wrong["Q"])); continue right = evaluate_row(ev, r.prompt, pl, pa, None, "stop") res.append((right["Q"], wrong["Q"])) rights = [a for a, b in res if a is not None] print(f"{ev:18s} n={len(res)} perfect-mean={sum(rights) / max(len(rights), 1):.2f} (n={len(rights)}, min={min(rights) if rights else None}) wrong-mean={sum(b for a, b in res) / len(res):.2f}", flush=True) for a, b in res: if a is not None and a < 9.99: bad += 1 print("imperfect gold cases:", bad)