src/data/token_stats.py

from __future__ import annotations

import argparse
from pathlib import Path

from src.data.io_utils import read_jsonl, write_csv
from src.data.normalize_text import word_count


def summarize_lengths(lengths: list[int]) -> dict[str, object]:
    lengths = sorted(lengths)
    if not lengths:
        return {
            "count": 0,
            "mean_words": 0,
            "p50_words": 0,
            "p95_words": 0,
            "p99_words": 0,
            "max_words": 0,
        }
    return {
        "count": len(lengths),
        "mean_words": round(sum(lengths) / len(lengths), 3),
        "p50_words": lengths[int(0.50 * (len(lengths) - 1))],
        "p95_words": lengths[int(0.95 * (len(lengths) - 1))],
        "p99_words": lengths[int(0.99 * (len(lengths) - 1))],
        "max_words": max(lengths),
    }


def add_rows(rows: list[dict], dataset: str, split: str, file_name: str, records: list[dict], fields: list[str]) -> None:
    for field in fields:
        lengths = [word_count(record.get(field)) for record in records if record.get(field) is not None]
        summary = summarize_lengths(lengths)
        rows.append({"dataset": dataset, "split": split, "file": file_name, "field": field, **summary})


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--data-root", type=Path, default=Path("data_processed"))
    parser.add_argument("--output", type=Path, default=Path("outputs/stats/token_length_report.csv"))
    args = parser.parse_args()

    rows: list[dict] = []
    for split in ["train", "dev", "test"]:
        path = args.data_root / "vifactcheck" / f"claims_{split}.jsonl"
        if path.exists():
            add_rows(rows, "vifactcheck", split, path.name, read_jsonl(path), ["claim", "context"])
    for file_name, field, split in [
        ("context_sentences.jsonl", "text", "all"),
        ("context_chunks.jsonl", "text", "all"),
        ("gold_evidence.jsonl", "text", "all"),
    ]:
        path = args.data_root / "vifactcheck" / file_name
        if path.exists():
            add_rows(rows, "vifactcheck", split, file_name, read_jsonl(path), [field])

    for split in ["train_inner", "dev_inner", "local_test", "hidden_test"]:
        path = args.data_root / "averitec" / f"claims_{split}.jsonl"
        if path.exists():
            add_rows(rows, "averitec", split, path.name, read_jsonl(path), ["claim"])
    for file_name in ["qa_evidence.jsonl", "evidence_store.jsonl"]:
        path = args.data_root / "averitec" / file_name
        if path.exists():
            fields = ["question", "answer", "evidence_text"] if file_name == "qa_evidence.jsonl" else ["text"]
            add_rows(rows, "averitec", "all_labeled_evidence", file_name, read_jsonl(path), fields)

    for split in ["train", "dev", "test"]:
        path = args.data_root / "healthver" / f"pairs_{split}.jsonl"
        if path.exists():
            add_rows(rows, "healthver", split, path.name, read_jsonl(path), ["claim", "evidence"])
        grouped_path = args.data_root / "healthver" / f"claims_grouped_{split}.jsonl"
        if grouped_path.exists():
            add_rows(rows, "healthver", split, grouped_path.name, read_jsonl(grouped_path), ["claim"])
    path = args.data_root / "healthver" / "evidence_sentences.jsonl"
    if path.exists():
        add_rows(rows, "healthver", "all", path.name, read_jsonl(path), ["text"])

    write_csv(args.output, rows)
    print(f"Wrote token length report to {args.output}")


if __name__ == "__main__":
    main()