scripts/aime_ensemble_selector.py

from __future__ import annotations

import argparse
import json
import re
from collections import Counter
from pathlib import Path

from eval_answer_utils import extract_final_answer


RUNS = {
    "qwen_full": "runs/20260408T070936Z_aime20_qwen4b_instruct_english/results.jsonl",
    "qwen_repair": "runs/20260408T070936Z_aime20_qwen4b_instruct_english_repair/results.jsonl",
    "qwen_cont": (
        "runs/20260408T070936Z_aime20_qwen4b_instruct_english_continuation_repaired/results.jsonl"
    ),
    "alt": "runs/20260408T070936Z_aime20_qwen4b_instruct_lfm25_alternating_repaired/results.jsonl",
    "gemma": "runs/20260408T103238Z_aime20_gemma4e4b_thinking/results.jsonl",
    "gemma_sr": "runs/20260409T125112Z_aime20_gemma4e4b_thinking_self_repair/results.jsonl",
}

PRIORITY = ["qwen_cont", "alt", "gemma", "gemma_sr", "qwen_full"]
REPAIR_BAD_MARKERS = (
    "known solution",
    "known problem",
    "known similar",
    "given the time",
    "not sure",
    "guess",
    "i think i will go",
    "in the absence",
    "typo",
)
INTEGER_ANSWER_RE = re.compile(r"^0*\d{1,3}$")


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser()
    parser.add_argument("--out-dir", required=True)
    return parser.parse_args()


def load_records(path: Path) -> list[dict[str, object]]:
    return [json.loads(line) for line in path.read_text().splitlines() if line.strip()]


def answer_source(record: dict[str, object]) -> str:
    return str(
        record.get("raw_output") or record.get("final") or record.get("predicted_answer") or ""
    )


def integer_answer(record: dict[str, object]) -> str | None:
    answer = extract_final_answer(answer_source(record)).strip().strip(" .")
    if INTEGER_ANSWER_RE.fullmatch(answer) is None:
        return None
    return str(int(answer))


def clean_repair(record: dict[str, object]) -> bool:
    source = answer_source(record).lower()
    return (
        integer_answer(record) is not None
        and record.get("finish_reason") == "stop"
        and not any(marker in source for marker in REPAIR_BAD_MARKERS)
    )


def majority_vote(votes: list[tuple[str, str]]) -> tuple[str | None, list[str]]:
    if not votes:
        return None, []
    counts = Counter(answer for answer, _ in votes)
    answer = sorted(
        counts,
        key=lambda value: (
            -counts[value],
            min(PRIORITY.index(name) for candidate, name in votes if candidate == value),
        ),
    )[0]
    return answer, [name for candidate, name in votes if candidate == answer]


def select_answer(records: dict[str, dict[str, object]]) -> tuple[str | None, list[str], str]:
    votes = []
    for name in PRIORITY:
        if name in records:
            answer = integer_answer(records[name])
            if answer is not None:
                votes.append((answer, name))

    answer, sources = majority_vote(votes)
    rationale = f"majority_{len(sources)}" if sources else "no_integer_vote"

    repair = records.get("qwen_repair")
    if repair is not None and clean_repair(repair) and len(sources) <= 2:
        repair_answer = integer_answer(repair)
        if repair_answer is None:
            raise ValueError(f"clean repair without integer answer for {repair['id']}")
        return repair_answer, ["qwen_repair"], "clean_qwen_repair_override"

    return answer, sources, rationale


def evaluate(records_by_id: dict[str, dict[str, dict[str, object]]]) -> list[dict[str, object]]:
    rows = []
    for problem_id in sorted(records_by_id):
        records = records_by_id[problem_id]
        gold = str(int(str(next(iter(records.values()))["gold_answer"])))
        answer, sources, rationale = select_answer(records)
        votes = {}
        for name, record in sorted(records.items()):
            vote = integer_answer(record)
            if vote is not None:
                votes[name] = vote
        rows.append(
            {
                "id": problem_id,
                "gold_answer": gold,
                "selected_answer": answer,
                "correct": answer == gold,
                "selected_sources": sources,
                "rationale": rationale,
                "integer_votes": votes,
                "oracle_correct_in_selected_runs": any(
                    bool(record.get("correct")) for record in records.values()
                ),
            }
        )
    return rows


def load_runs() -> dict[str, dict[str, dict[str, object]]]:
    records_by_id: dict[str, dict[str, dict[str, object]]] = {}
    for name, path_text in RUNS.items():
        path = Path(path_text)
        if not path.exists():
            raise FileNotFoundError(path)
        for record in load_records(path):
            problem_id = str(record["id"])
            if problem_id.startswith("aime24_"):
                records_by_id.setdefault(problem_id, {})[name] = record
    return records_by_id


def summarize(
    rows: list[dict[str, object]],
    records_by_id: dict[str, dict[str, dict[str, object]]],
) -> dict[str, object]:
    total = len(rows)
    correct = sum(bool(row["correct"]) for row in rows)
    oracle_correct = sum(bool(row["oracle_correct_in_selected_runs"]) for row in rows)
    single_run_scores = {
        name: sum(
            bool(records.get(name, {}).get("correct")) for records in records_by_id.values()
        )
        for name in RUNS
    }
    return {
        "correct": correct,
        "total": total,
        "accuracy": correct / total,
        "oracle_correct_in_selected_runs": oracle_correct,
        "best_single_run_correct": max(single_run_scores.values()),
        "single_run_scores": single_run_scores,
        "runs": RUNS,
    }


def main() -> None:
    args = parse_args()
    records_by_id = load_runs()
    rows = evaluate(records_by_id)
    summary = summarize(rows, records_by_id)

    out_dir = Path(args.out_dir)
    out_dir.mkdir(parents=True, exist_ok=True)
    with (out_dir / "results.jsonl").open("w") as handle:
        for row in rows:
            handle.write(json.dumps(row, ensure_ascii=False) + "\n")
    (out_dir / "summary.json").write_text(
        json.dumps(summary, indent=2, ensure_ascii=False) + "\n"
    )
    print(json.dumps(summary, indent=2, ensure_ascii=False))


if __name__ == "__main__":
    main()