evaluate.py

#!/usr/bin/env python3
"""Score model answers for the trace-help dataset.

The parsing and matching rules below were fixed before any model was run and
have not been changed since. Grading is blind: `judge` receives only the
accepted answers and the answer text, never the setting the answer came from.

Usage:
    evaluate.py data/test.jsonl answers.jsonl

`answers.jsonl` — one JSON object per line:
    {"id": "orders-1", "setting": "with_trace", "repeat": 1, "output": "..."}
`setting` is either "no_trace" or "with_trace".
"""
from __future__ import annotations

import json
import random
import re
import statistics
import sys
from pathlib import Path

ABSTAIN = ("не знаю", "неизвестно", "нельзя определить", "недостаточно",
           "i don't know", "unknown", "cannot be determined")
NUMBER = re.compile(r"-?\d+")
SETTINGS = ("no_trace", "with_trace")


def tidy(text: str) -> str:
    text = text.strip().strip("«»\"'`").strip().rstrip(".")
    return " ".join(text.lower().split())


def extract(raw: str) -> str | None:
    """The last line containing ОТВЕТ: — everything after the colon."""

    found = None
    for line in raw.splitlines():
        if "ОТВЕТ:" in line.upper() or "ответ:" in line:
            found = line.split(":", 1)[1] if ":" in line else line
    return found


def judge(accepted: list[str], raw: str) -> str:
    """correct / abstain / wrong. Sees only accepted answers and the text."""

    said = extract(raw)
    if said is None:
        return "wrong"
    said = tidy(said)
    if any(word in said for word in ABSTAIN):
        return "abstain"
    for want in accepted:
        want = tidy(want)
        if said == want:
            return "correct"
        if want in ("да", "нет") and said.startswith(want):
            return "correct"
        if NUMBER.fullmatch(want) and NUMBER.findall(said) == [want]:
            return "correct"
    return "wrong"


def bootstrap(groups: list[list[tuple[str, str]]], rounds: int = 10000) -> dict:
    """95% interval, resampling whole programs with replacement."""

    rnd = random.Random(20260830)
    n = len(groups)
    keep = {k: [] for k in ("no_trace", "with_trace", "difference",
                            "wrong_no_trace", "wrong_with_trace", "wrong_difference")}
    for _ in range(rounds):
        flat = [item for _ in range(n) for item in groups[rnd.randrange(n)]]
        for setting in SETTINGS:
            mine = [v for s, v in flat if s == setting]
            keep[setting].append(sum(v == "correct" for v in mine) / len(mine))
            keep["wrong_" + setting].append(
                sum(v == "wrong" for v in mine) / len(mine))
        keep["difference"].append(keep["with_trace"][-1] - keep["no_trace"][-1])
        keep["wrong_difference"].append(
            keep["wrong_with_trace"][-1] - keep["wrong_no_trace"][-1])
    out = {}
    for key, values in keep.items():
        values.sort()
        out[key] = (values[int(0.025 * rounds)], values[int(0.975 * rounds)])
    return out


def pct(x: float) -> str:
    return f"{100 * x:.1f} %"


def main() -> int:
    rows = [json.loads(l) for l in Path(sys.argv[1]).read_text(
        encoding="utf-8").splitlines() if l.strip()]
    answers = [json.loads(l) for l in Path(sys.argv[2]).read_text(
        encoding="utf-8").splitlines() if l.strip()]

    accepted = {r["id"]: r["answers"] for r in rows}
    kind = {r["id"]: r["question_kind"] for r in rows}
    language = {r["id"]: r["language"] for r in rows}
    program = {r["id"]: r["program"] for r in rows}

    # --- blind part: only question id and answer text go in -----------------
    blind = [(i, a["id"], a["output"]) for i, a in enumerate(answers)]
    random.Random(1).shuffle(blind)
    verdicts = {i: judge(accepted[qid], text) for i, qid, text in blind}
    # --- from here on the setting may be looked at --------------------------
    for i, a in enumerate(answers):
        a["verdict"] = verdicts[i]

    def share(rows_, verdict):
        return sum(r["verdict"] == verdict for r in rows_) / len(rows_) if rows_ else 0.0

    print(f"answers: {len(answers)}; questions: {len(accepted)}; "
          f"programs: {len(set(program.values()))}; "
          f"repeats: {len({a.get('repeat', 1) for a in answers})}")
    print()
    print("| setting | correct | confidently wrong | abstained | answers |")
    print("|---|---|---|---|---|")
    for setting in SETTINGS:
        mine = [a for a in answers if a["setting"] == setting]
        print(f"| {setting} | {pct(share(mine, 'correct'))} | "
              f"{pct(share(mine, 'wrong'))} | {pct(share(mine, 'abstain'))} | "
              f"{len(mine)} |")

    groups = []
    for prog in sorted(set(program.values())):
        ids = {q for q, p in program.items() if p == prog}
        groups.append([(a["setting"], a["verdict"]) for a in answers
                       if a["id"] in ids])
    band = bootstrap(groups)
    print()
    print("95% intervals (10 000 resamples over programs):")
    print(f"  correct, no trace    {pct(band['no_trace'][0])} … {pct(band['no_trace'][1])}")
    print(f"  correct, with trace  {pct(band['with_trace'][0])} … {pct(band['with_trace'][1])}")
    print(f"  difference           {pct(band['difference'][0])} … {pct(band['difference'][1])}")
    print(f"  confidently wrong, difference {pct(band['wrong_difference'][0])} … "
          f"{pct(band['wrong_difference'][1])}")

    repeats = sorted({a.get("repeat", 1) for a in answers})
    if len(repeats) > 1:
        print()
        print("| repeat | no trace | with trace |")
        print("|---|---|---|")
        per = {}
        for rep in repeats:
            line = [share([a for a in answers if a.get("repeat", 1) == rep
                           and a["setting"] == s], "correct") for s in SETTINGS]
            per[rep] = line
            print(f"| {rep} | {pct(line[0])} | {pct(line[1])} |")
        for idx, title in enumerate(SETTINGS):
            vals = [v[idx] for v in per.values()]
            print(f"  {title}: range {pct(min(vals))} … {pct(max(vals))}, "
                  f"standard deviation {pct(statistics.stdev(vals))}")

    for title, table in (("language", language), ("question kind", kind)):
        print()
        print(f"| {title} | correct, no trace | correct, with trace | answers each |")
        print("|---|---|---|---|")
        for value in sorted(set(table.values())):
            ids = {q for q, v in table.items() if v == value}
            line = [[a for a in answers if a["id"] in ids and a["setting"] == s]
                    for s in SETTINGS]
            print(f"| {value} | {pct(share(line[0], 'correct'))} | "
                  f"{pct(share(line[1], 'correct'))} | {len(line[0])} |")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())