scripts/build_hf_repo.py

import argparse
import json
import shutil
import zipfile
from pathlib import Path

import pandas as pd
import pyarrow as pa
import pyarrow.parquet as pq


RAW_FILES = {
    "train": "FunQA_train.json",
    "validation": "FunQA_val.json",
    "test": "FunQA_test.json",
    "mcqa_test": "Funqa_mcqa_v1.json",
}

VIDEO_ARCHIVES = {
    "validation": "val.zip",
    "test": "test.zip",
    "train": "train.zip",
}


def load_raw_rows(raw_dir: Path, split: str):
    with (raw_dir / RAW_FILES[split]).open("r", encoding="utf-8") as f:
        rows = json.load(f)
    if split == "validation":
        missing_videos = {"C_KT_6_6347_6427.mp4"}
        rows = [row for row in rows if row.get("visual_input") not in missing_videos]
    return rows


def write_parquet(rows, output_path: Path):
    df = pd.DataFrame(rows)
    output_path.parent.mkdir(parents=True, exist_ok=True)

    if "output" in df.columns:
        schema = pa.schema(
            [
                ("instruction", pa.string()),
                ("visual_input", pa.string()),
                ("output", pa.string()),
                ("task", pa.string()),
            ]
        )
        df = df[["instruction", "visual_input", "output", "task"]]
    else:
        schema = pa.schema(
            [
                ("instruction", pa.string()),
                ("visual_input", pa.string()),
                ("gt", pa.string()),
                ("id", pa.string()),
            ]
        )
        df = df[["instruction", "visual_input", "gt", "id"]]

    table = pa.Table.from_pandas(df, schema=schema, preserve_index=False)
    pq.write_table(table, output_path)


def move_raw_files(repo_root: Path, raw_dir: Path):
    raw_dir.mkdir(parents=True, exist_ok=True)
    for filename in list(RAW_FILES.values()) + list(VIDEO_ARCHIVES.values()):
        src = repo_root / filename
        dst = raw_dir / filename
        if src.exists() and not dst.exists():
            shutil.move(str(src), str(dst))


def extract_split_videos(raw_dir: Path, videos_dir: Path, split: str):
    archive_name = VIDEO_ARCHIVES[split]
    archive_path = raw_dir / archive_name
    if not archive_path.exists():
        return
    videos_dir.mkdir(parents=True, exist_ok=True)
    marker = videos_dir / f".{split}_extracted"
    if marker.exists():
        return
    with zipfile.ZipFile(archive_path) as zf:
        zf.extractall(videos_dir)
    marker.write_text("ok\n", encoding="utf-8")


def normalize_video_layout(videos_dir: Path, split: str):
    alias = {"validation": "val"}.get(split, split)
    alias_root = videos_dir / alias
    expected_root = videos_dir / split
    if alias_root.exists() and alias_root != expected_root and not expected_root.exists():
        shutil.move(str(alias_root), str(expected_root))

    legacy_root = videos_dir / split / split
    if legacy_root.exists():
        expected_root.mkdir(parents=True, exist_ok=True)
        for child in legacy_root.iterdir():
            target = expected_root / child.name
            if not target.exists():
                shutil.move(str(child), str(target))

    legacy_alias_root = videos_dir / split / alias
    if legacy_alias_root.exists():
        expected_root.mkdir(parents=True, exist_ok=True)
        for child in legacy_alias_root.iterdir():
            target = expected_root / child.name
            if not target.exists():
                shutil.move(str(child), str(target))


def ensure_layout(repo_root: Path):
    move_raw_files(repo_root, repo_root / "raw")
    for split in ["test", "validation", "train"]:
        extract_split_videos(repo_root / "raw", repo_root / "videos", split)
        normalize_video_layout(repo_root / "videos", split)


def main():
    parser = argparse.ArgumentParser(description="Build a HF Hub-style FunQA dataset repo with original columns.")
    parser.add_argument("--repo-root", type=Path, default=Path("."))
    args = parser.parse_args()

    repo_root = args.repo_root.resolve()
    ensure_layout(repo_root)

    raw_dir = repo_root / "raw"
    data_dir = repo_root / "data"

    for split in RAW_FILES:
        rows = load_raw_rows(raw_dir, split)
        write_parquet(rows, data_dir / f"{split}.parquet")

    print(f"Built parquet splits under: {data_dir}")


if __name__ == "__main__":
    main()