score.py

"""Reference scoring implementation for asmachta.json.

Verifies that a model's claimed quotes actually appear in the source
document -- the same method used to produce this dataset's own
`verified` / `verification_method` fields (see README, "Support
sentences: how grounding works"). No dependencies beyond the standard
library.

Expected model output format, one claim per line:

    1. <claim text> [<verbatim quote from source_text>]
    2. <claim text> [<verbatim quote from source_text>]

Usage as a library:

    from score import parse_claims, score_answer

    claims = parse_claims(model_output_text)
    result = score_answer(claims, record["source_text"])
    print(result["precision"], result["claims"])

Usage as a script (demo against the dataset's own reference answers,
which should score at or near 100% grounded since they are the source
of the verified attribution spans):

    python3 score.py asmachta.json
"""
from __future__ import annotations

import difflib
import json
import re
import sys

CLAIM_RE = re.compile(r"^\s*\d+\.\s*(.+?)\s*\[(.+?)\]\s*$", re.MULTILINE)

# Fallback for output where the model prefixed the numbered claim with
# leftover text on the same line (e.g. reasoning-model self-talk, or a
# stray closing </think> tag never on its own line) -- CLAIM_RE's line
# anchor then finds nothing even though a real, checkable quote is
# right there. This pattern doesn't require the claim to start at the
# beginning of a line, only that "N." is immediately followed by claim
# text and a bracketed quote. Only used as a fallback: if CLAIM_RE finds
# anything at all, its output is used unchanged, so this never changes
# scoring for output that was already parsing correctly.
CLAIM_RE_FALLBACK = re.compile(r"(?:^|\.)?\s*\d+\.\s*(.+?)\s*\[([^\[\]]{3,})\]", re.MULTILINE)

# Same 0.94 similarity threshold this dataset's own offset-validation
# pipeline used for near-exact (whitespace/punctuation-level) matches.
FUZZY_THRESHOLD = 0.94


def parse_claims(text: str) -> list[dict]:
    """Extract (claim, quote) pairs from `N. claim [quote]`-formatted text.

    Tries the strict, line-anchored pattern first; only falls back to a
    more permissive pattern if that finds nothing at all. This recovers
    real quotes hidden behind formatting noise without changing the
    result for output that already parses cleanly -- see CLAIM_RE_FALLBACK.
    """
    matches = CLAIM_RE.findall(text)
    if not matches:
        matches = CLAIM_RE_FALLBACK.findall(text)
    return [{"text": m[0].strip(), "quote": m[1].strip()} for m in matches]


def _normalize(s: str) -> str:
    return re.sub(r"\s+", " ", s).strip()


def verify_quote(quote: str, source_text: str) -> dict:
    """Check whether `quote` appears in `source_text`, exact/normalized/fuzzy."""
    if quote in source_text:
        return {"verified": True, "method": "exact", "score": 1.0}

    if _normalize(quote) in _normalize(source_text):
        return {"verified": True, "method": "normalized_space", "score": 1.0}

    sm = difflib.SequenceMatcher(None, source_text, quote, autojunk=False)
    match = sm.find_longest_match(0, len(source_text), 0, len(quote))
    if match.size == 0:
        return {"verified": False, "method": "fuzzy", "score": 0.0}
    pad = len(quote) - match.size
    window = source_text[max(0, match.a - pad): min(len(source_text), match.a + match.size + pad)]
    ratio = difflib.SequenceMatcher(None, _normalize(window), _normalize(quote)).ratio()
    return {"verified": ratio >= FUZZY_THRESHOLD, "method": "fuzzy", "score": ratio}


def score_answer(claims: list[dict], source_text: str) -> dict:
    """Score a list of {"text", "quote"} claims against a source document.

    Returns per-claim verification results plus overall attribution
    precision (grounded claims / total claims).
    """
    scored = []
    grounded = 0
    for c in claims:
        v = verify_quote(c["quote"], source_text)
        scored.append({**c, **v})
        if v["verified"]:
            grounded += 1
    n = len(claims)
    return {
        "n_claims": n,
        "n_grounded": grounded,
        "precision": grounded / n if n else None,
        "claims": scored,
    }


def score_model_output(text: str, source_text: str) -> dict:
    """End-to-end: parse a raw model response and score it against a document."""
    return score_answer(parse_claims(text), source_text)


def _demo(dataset_path: str) -> None:
    with open(dataset_path, encoding="utf-8") as f:
        records = json.load(f)

    # `reference_answer` is plain prose -- the dataset's own claim/quote
    # structure lives in `claims[].attribution[]`, not embedded as
    # "N. claim [quote]" text. Re-verify verify_quote() against that
    # structure directly, for the first 5 answerable records: this
    # should reproduce every `verified: true` the dataset already ships.
    print("Re-verifying this file's own claims[].attribution[] spans with")
    print("verify_quote() (sanity check -- should match every 'verified' field):\n")

    shown = 0
    for r in records:
        if r["difficulty"] == 0:
            continue
        n = sum(len(c["attribution"]) for c in r["claims"])
        agree = 0
        for c in r["claims"]:
            for a in c["attribution"]:
                v = verify_quote(a["source_excerpt"], r["source_text"])
                if v["verified"] == a["verified"]:
                    agree += 1
        print(f"  {r['id']}: {agree}/{n} spans match the dataset's own 'verified' label")
        shown += 1
        if shown >= 5:
            break

    print("\nScoring a MODEL's raw output (the 'N. claim [quote]' format from")
    print("the README) against a source document -- a fabricated example, one")
    print("real claim and one hallucinated claim:")
    r = next(r for r in records if r["difficulty"] != 0)
    real_excerpt = r["claims"][0]["attribution"][0]["source_excerpt"]
    fake_output = (
        f"1. טענה אמיתית שנתמכת במסמך. [{real_excerpt}]\n"
        f"2. משהו שלא נכתב במסמך כלל ולעולם לא יימצא שם. [ציטוט מומצא שלא קיים]"
    )
    result = score_model_output(fake_output, r["source_text"])
    print(f"  {result['n_grounded']}/{result['n_claims']} claims grounded "
          f"(precision={result['precision']:.2f})")


if __name__ == "__main__":
    if len(sys.argv) != 2:
        print(f"Usage: python3 {sys.argv[0]} asmachta.json", file=sys.stderr)
        sys.exit(1)
    _demo(sys.argv[1])