scripts/export_hf.py

#!/usr/bin/env python3
"""Package verse and benchmark JSONL files for the Hugging Face folder."""

from __future__ import annotations

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

ROOT = Path(__file__).resolve().parents[2]
HF = Path(__file__).resolve().parents[1]
DATA = ROOT / "data"
OUT = HF / "data"

BAHAR = re.compile(r"\s*\([^)]*\)\s*$")


def arkans(rhythm: str) -> str:
    return BAHAR.sub("", rhythm or "").strip()


def leaked_keys(probe_path: Path, threshold: float) -> set[tuple]:
    best: dict[tuple, float] = {}
    with probe_path.open(encoding="utf-8") as handle:
        for line in handle:
            row = json.loads(line)
            key = (row["poem_id"], row["couplet_index"])
            score = float(row.get("score") or 0)
            if score > best.get(key, 0):
                best[key] = score
    return {key for key, score in best.items() if score > threshold}


def export_verses(src: Path, dest: Path) -> dict:
    dest.parent.mkdir(parents=True, exist_ok=True)
    n = 0
    metres = Counter()
    with src.open(encoding="utf-8") as inp, dest.open("w", encoding="utf-8") as out:
        for line in inp:
            row = json.loads(line)
            rhythm = arkans(row["rhythm"])
            if not row.get("verse") or not rhythm:
                continue
            record = {
                "verse": row["verse"],
                "rhythm": rhythm,
                "poem_id": row.get("poem_id"),
                "poet_id": row.get("poet_id"),
                "couplet_index": row.get("couplet_index"),
                "mesra_index": row.get("verse_order"),
            }
            out.write(json.dumps(record, ensure_ascii=False))
            out.write("\n")
            metres[rhythm] += 1
            n += 1
    return {"rows": n, "metres": len(metres)}


def export_benchmark(src: Path, dest: Path, drop: set[tuple]) -> dict:
    dest.parent.mkdir(parents=True, exist_ok=True)
    n = 0
    dropped = 0
    metres = Counter()
    with src.open(encoding="utf-8") as inp, dest.open("w", encoding="utf-8") as out:
        for line in inp:
            row = json.loads(line)
            key = (row["poem_id"], row["couplet_index"])
            if key in drop:
                dropped += 1
                continue
            record = {
                "first": row["first"],
                "second": row["second"],
                "rhythm": arkans(row["rhythm"]),
                "poet": row.get("poet"),
                "poet_id": row.get("poet_id"),
                "book": row.get("book"),
                "poem_id": row.get("poem_id"),
                "couplet_index": row.get("couplet_index"),
            }
            out.write(json.dumps(record, ensure_ascii=False))
            out.write("\n")
            metres[record["rhythm"]] += 1
            n += 1
    return {"rows": n, "dropped": dropped, "metres": len(metres), "metre_counts": metres}


def main() -> None:
    threshold = 0.6
    drop = leaked_keys(DATA / "leakage_probe.jsonl", threshold)
    verses = export_verses(DATA / "verses.jsonl", OUT / "verses.jsonl")
    bench = export_benchmark(DATA / "leakage_pool.jsonl", OUT / "benchmark.jsonl", drop)
    print(json.dumps({
        "verses": verses,
        "benchmark": {
            "rows": bench["rows"],
            "dropped_similarity_gt_0.6": bench["dropped"],
            "metres": bench["metres"],
        },
    }, ensure_ascii=False, indent=2))
    print("--- benchmark metres ---")
    for rhythm, count in bench["metre_counts"].most_common():
        print(f"  {count:5d}  {rhythm}")


if __name__ == "__main__":
    main()