legex/evaluation/scoring.py

"""Compare Goldenset gold labels against inference output.

Reads each country's gold labels — from ``Goldenset_*.xlsx`` (the maintainers'
workbooks) or, with ``--gold-dir``, from the published goldenset JSONL — and
the matching predictions JSONL, then reports per-field agreement.
"""

import argparse
import logging
import sys
from pathlib import Path

from openpyxl import load_workbook

from legex import published
from legex.evaluation.comparison import (
    BUCKETS,
    classify_cell,
    derived,
    is_label_column,
    normalise,
)
from legex.inference import inference_output_path
from legex.utils import (
    countries_with_goldenset,
    goldenset_path,
    goldenset_sheet,
    read_inference_jsonl,
)

log = logging.getLogger(__name__)


def _read_goldenset_rows(cc: str) -> tuple[list[str], dict[str, dict[str, str]]]:
    """Return (label_columns, rows_by_case_id) for a country's Goldenset."""
    wb = load_workbook(goldenset_path(cc), read_only=True, data_only=True)
    ws = goldenset_sheet(wb)
    rows = ws.iter_rows(values_only=True)
    header = [str(c) if c is not None else "" for c in next(rows)]
    if "case_id" not in header:
        raise ValueError(f"{goldenset_path(cc)} GOLDENSET sheet missing case_id column")
    label_columns = [h for h in header if is_label_column(h)]

    by_id: dict[str, dict[str, str]] = {}
    for row in rows:
        if not any(row):
            continue
        cells = dict(zip(header, row))
        case_id = normalise(cells.get("case_id"))
        if not case_id:
            continue
        labels = {col: normalise(cells.get(col)) for col in label_columns}
        if not any(labels.values()):
            continue
        by_id[case_id] = labels
    return label_columns, by_id


def _read_predictions(path: Path) -> dict[str, dict[str, str]]:
    by_id: dict[str, dict[str, str]] = {}
    for row in read_inference_jsonl(path):
        case_id = normalise(row.get("case_id"))
        if not case_id:
            continue
        # Rows whose inference failed are treated like absent inference
        if normalise(row.get("error")):
            continue
        labels = {k: normalise(v) for k, v in row.items() if is_label_column(k)}
        if not any(labels.values()):
            continue
        by_id[case_id] = {k: normalise(v) for k, v in row.items()}
    return by_id


def _coverage(
    gold: dict[str, dict[str, str]], preds: dict[str, dict[str, str]]
) -> dict[str, int]:
    gold_ids, pred_ids = set(gold), set(preds)
    return {
        "gold": len(gold_ids),
        "pred": len(pred_ids),
        "overlap": len(gold_ids & pred_ids),
        "missing": len(gold_ids - pred_ids),
        "extra": len(pred_ids - gold_ids),
    }


def score_country(
    cc: str,
    prompt_version: str,
    source: str,
    model: str,
    verbose: bool = True,
    *,
    gold_dir: Path | None = None,
    inference_dir: Path | None = None,
) -> tuple[dict[str, dict[str, int]], dict[str, int]] | None:
    """Return (per-column counters, case coverage stats).

    ``gold_dir`` / ``inference_dir`` switch the respective input to the
    published JSONL bundles (see ``legex.published``); by default the
    maintainers' XLSX workbooks and working inference files are read.
    """
    pred_path = (
        published.inference_file(inference_dir, cc, model)
        if inference_dir is not None
        else inference_output_path(cc, prompt_version, source, model)
    )
    if not pred_path.exists():
        if verbose:
            log.warning(f"[{cc}] missing predictions {pred_path}, skipping")
        return None
    gold_path = published.gold_file(gold_dir, cc) if gold_dir is not None else goldenset_path(cc)
    if not gold_path.exists():
        if verbose:
            log.warning(f"[{cc}] missing goldenset {gold_path}, skipping")
        return None

    if gold_dir is not None:
        label_columns, gold = published.load_gold_labels(gold_dir, cc)
    else:
        label_columns, gold = _read_goldenset_rows(cc)
    preds = _read_predictions(pred_path)
    stats = _coverage(gold, preds)

    counters: dict[str, dict[str, int]] = {col: {b: 0 for b in BUCKETS} for col in label_columns}
    if verbose:
        print()
        log.info(
            f"[{cc}] gold={stats['gold']} pred={stats['pred']} overlap={stats['overlap']} "
            f"missing={stats['missing']} extra={stats['extra']}"
        )

    for case_id in (cid for cid in gold if cid in preds):
        g, p = gold[case_id], preds[case_id]
        for col in label_columns:
            counters[col][classify_cell(g.get(col, ""), p.get(col, ""), col)] += 1
    return counters, stats


def _print_report(
    cc: str, counters: dict[str, dict[str, int]], coverage: dict[str, int] | None = None
) -> None:
    name_width = max(max((len(c) for c in counters), default=0), len("column"))
    print(f"=== {cc} ===")
    if coverage is not None:
        print(
            f"cases: gold={coverage['gold']} pred={coverage['pred']} "
            f"overlap={coverage['overlap']} missing={coverage['missing']} "
            f"extra={coverage['extra']}"
        )
    print(
        f"{'column'.ljust(name_width)}  "
        f"{'TP':>5}  {'Mism':>5}  {'Miss':>5}  {'Hallu':>5}  {'TN':>5}  "
        f"{'P':>7}  {'R':>7}  {'F1':>5}"
    )
    for col, c in counters.items():
        p, r, f1 = derived(c)
        f1_s = f"{f1:.2f}" if (p + r) else "   - "
        print(
            f"{col.ljust(name_width)}  "
            f"{c['tp']:>5}  {c['mismatch']:>5}  {c['missed']:>5}  "
            f"{c['hallucinated']:>5}  {c['tn']:>5}  "
            f"{p:>7.2%}  {r:>7.2%}  {f1_s:>5}"
        )


def evaluate(
    countries: list[str] | None,
    model: str,
    prompt_version: str,
    source: str,
    gold_dir: Path | None = None,
    inference_dir: Path | None = None,
) -> None:
    overall: dict[str, dict[str, int]] = {
        col: {b: 0 for b in BUCKETS} for col in published.LABEL_FIELDS
    }
    overall_coverage = {"gold": 0, "pred": 0, "overlap": 0, "missing": 0, "extra": 0}

    if countries is None:
        countries = (
            published.countries_with_gold(gold_dir)
            if gold_dir is not None
            else countries_with_goldenset()
        )
    seen_any = False
    for cc in countries:
        result = score_country(
            cc, prompt_version, source, model,
            gold_dir=gold_dir, inference_dir=inference_dir,
        )
        if result is None:
            continue
        counters, coverage = result
        seen_any = True
        _print_report(cc, counters, coverage)
        for key in overall_coverage:
            overall_coverage[key] += coverage[key]
        for col, c in counters.items():
            overall.setdefault(col, {b: 0 for b in BUCKETS})
            for b in BUCKETS:
                overall[col][b] += c[b]

    if seen_any:
        _print_report("ALL", overall, overall_coverage)


def main() -> None:
    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s [%(levelname)s] %(message)s",
        handlers=[logging.StreamHandler(sys.stderr)],
    )
    parser = argparse.ArgumentParser(
        prog="legex-evaluate",
        description="Compare Goldenset labels against LLM inference output.",
    )
    parser.add_argument(
        "--country", action="extend", nargs="+", dest="countries",
        help="Country code(s). Repeatable and/or space-separated. Defaults to all with a Goldenset.",
    )
    parser.add_argument(
        "--model", required=True,
        help="Model id of the predictions to evaluate (must match the classify run).",
    )
    parser.add_argument("--prompt_version", default="v3", help="Prompt version (default: v3).")
    parser.add_argument(
        "--gold-dir", type=Path, default=None,
        help="Read gold labels from published goldenset JSONL under this directory "
             "instead of the XLSX workbooks.",
    )
    parser.add_argument(
        "--inference-dir", type=Path, default=None,
        help="Read predictions from published inference JSONL under this directory "
             "instead of the working files.",
    )
    source = parser.add_mutually_exclusive_group()
    source.add_argument("--full_text", dest="source", action="store_const", const="full_text",
                        help="Evaluate predictions made from the full_text column.")
    source.add_argument("--pdf", dest="source", action="store_const", const="pdf",
                        help="Evaluate predictions made from PDFs.")
    args = parser.parse_args()
    if args.inference_dir is None and args.source is None:
        parser.error("one of --full_text / --pdf is required (unless --inference-dir is used)")
    evaluate(
        countries=args.countries, model=args.model,
        prompt_version=args.prompt_version, source=args.source,
        gold_dir=args.gold_dir, inference_dir=args.inference_dir,
    )


if __name__ == "__main__":
    main()