scripts/prepare_three_benchmarks.py

#!/usr/bin/env python3
"""Prepare pinned CV-Bench, BLINK-val, and VStar data as auditable JSONL manifests."""

from __future__ import annotations

import argparse
import hashlib
import io
import json
import re
from collections import Counter
from pathlib import Path
from typing import Any, Iterable

import pyarrow.parquet as pq
from PIL import Image


REVISIONS = {
    "cvbench": "bc284db50d036958861cb60cdd7b77612052ce0d",
    "blink": "a3666eb249237ba3d5eca8db21176cc47967e040",
    "vstar": "d9ae62c903da0c98336e85c5ee89cd863b04b4da",
}
IMAGE_EXTENSIONS = {
    "BMP": ".bmp",
    "GIF": ".gif",
    "JPEG": ".jpg",
    "PNG": ".png",
    "TIFF": ".tiff",
    "WEBP": ".webp",
}
OPTION_RE = re.compile(r"^\s*\(?([A-Z])\)?\s*$", re.IGNORECASE)


def _sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for chunk in iter(lambda: handle.read(8 * 1024 * 1024), b""):
            digest.update(chunk)
    return digest.hexdigest()


def _write_json(path: Path, value: Any) -> None:
    path.write_text(
        json.dumps(value, ensure_ascii=False, indent=2, sort_keys=True) + "\n",
        encoding="utf-8",
    )


def _write_manifest(output_dir: Path, records: list[dict], metadata: dict) -> None:
    output_dir.mkdir(parents=True, exist_ok=True)
    manifest = output_dir / "manifest.jsonl"
    with manifest.open("w", encoding="utf-8") as handle:
        for record in records:
            handle.write(json.dumps(record, ensure_ascii=False) + "\n")
    metadata = {
        **metadata,
        "samples": len(records),
        "manifest_sha256": _sha256(manifest),
        "config_counts": dict(sorted(Counter(row["config"] for row in records).items())),
        "image_count": sum(len(row["images"]) for row in records),
    }
    _write_json(output_dir / "metadata.json", metadata)


def _expected(answer: Any, choices: list[Any]) -> str:
    match = OPTION_RE.fullmatch(str(answer))
    if not match:
        raise ValueError(f"Answer is not an option letter: {answer!r}")
    letter = match.group(1).upper()
    if ord(letter) - ord("A") >= len(choices):
        raise ValueError(f"Answer {letter} is outside {len(choices)} choices")
    return letter


def _image_extension(data: bytes, suggested: str | None = None) -> str:
    if suggested:
        suffix = Path(suggested).suffix.lower()
        if suffix in {".jpg", ".jpeg", ".png", ".webp", ".bmp", ".gif", ".tif", ".tiff"}:
            return ".jpg" if suffix == ".jpeg" else suffix
    with Image.open(io.BytesIO(data)) as image:
        return IMAGE_EXTENSIONS.get(str(image.format).upper(), ".img")


def _save_embedded_image(value: dict, output_base: Path, stem: str) -> tuple[str, int, int]:
    data = value.get("bytes")
    if not data:
        raise ValueError(f"Embedded image has no bytes: {value!r}")
    extension = _image_extension(data, value.get("path"))
    path = output_base / f"{stem}{extension}"
    path.parent.mkdir(parents=True, exist_ok=True)
    if not path.exists():
        path.write_bytes(data)
    elif path.read_bytes() != data:
        raise RuntimeError(f"Refusing to overwrite non-identical image: {path}")
    with Image.open(io.BytesIO(data)) as image:
        width, height = image.size
        image.verify()
    return path.name, width, height


def _iter_parquet(path: Path) -> Iterable[tuple[int, dict]]:
    row_index = 0
    parquet = pq.ParquetFile(path)
    for batch in parquet.iter_batches(batch_size=32):
        for row in batch.to_pylist():
            yield row_index, row
            row_index += 1


def prepare_cvbench(raw: Path, output: Path) -> None:
    records: list[dict] = []
    image_dir = output / "images"
    sources = [("2D", raw / "test_2d.parquet"), ("3D", raw / "test_3d.parquet")]
    for config, parquet in sources:
        if not parquet.is_file():
            raise FileNotFoundError(parquet)
        for row_index, row in _iter_parquet(parquet):
            sample_id = f"{config}-{int(row['idx']):04d}"
            filename, width, height = _save_embedded_image(
                row["image"], image_dir, sample_id
            )
            choices = [str(choice) for choice in row["choices"]]
            _expected(row["answer"], choices)
            records.append(
                {
                    "sample_id": sample_id,
                    "benchmark": "cvbench",
                    "split": "test",
                    "config": config,
                    "task": str(row["task"]),
                    "row_index": len(records),
                    "source_row_index": row_index,
                    "images": [f"images/{filename}"],
                    "image_sizes": [[width, height]],
                    "question": str(row["question"]),
                    "choices": choices,
                    "answer": str(row["answer"]),
                    "prompt": str(row["prompt"]),
                    "source_idx": int(row["idx"]),
                    "source_filename": row.get("filename"),
                }
            )
    if len(records) != 2638 or Counter(row["config"] for row in records) != Counter({"2D": 1438, "3D": 1200}):
        raise RuntimeError("CV-Bench count contract failed")
    _write_manifest(
        output,
        records,
        {
            "dataset": "nyu-visionx/CV-Bench",
            "revision": REVISIONS["cvbench"],
            "split": "test",
            "scope": "full official test",
        },
    )


def prepare_blink(raw: Path, output: Path) -> None:
    records: list[dict] = []
    image_dir = output / "images"
    parquets = sorted(raw.glob("*/val-*.parquet"))
    if len(parquets) != 14:
        raise RuntimeError(f"Expected 14 BLINK val parquet files, got {len(parquets)}")
    for parquet in parquets:
        config = parquet.parent.name
        for source_row_index, row in _iter_parquet(parquet):
            sample_id = str(row["idx"])
            image_paths: list[str] = []
            image_sizes: list[list[int]] = []
            for image_number in range(1, 5):
                embedded = row.get(f"image_{image_number}")
                if embedded is None:
                    continue
                filename, width, height = _save_embedded_image(
                    embedded,
                    image_dir,
                    f"{config}__{sample_id}__{image_number}",
                )
                image_paths.append(f"images/{filename}")
                image_sizes.append([width, height])
            if not image_paths:
                raise RuntimeError(f"BLINK sample has no images: {sample_id}")
            choices = [str(choice) for choice in row["choices"]]
            _expected(row["answer"], choices)
            records.append(
                {
                    "sample_id": sample_id,
                    "benchmark": "blink",
                    "split": "val",
                    "config": config,
                    "task": str(row["sub_task"]),
                    "row_index": len(records),
                    "source_row_index": source_row_index,
                    "images": image_paths,
                    "image_sizes": image_sizes,
                    "question": str(row["question"]),
                    "choices": choices,
                    "answer": str(row["answer"]),
                    "prompt": str(row["prompt"]),
                }
            )
    expected_counts = {
        "Art_Style": 117,
        "Counting": 120,
        "Forensic_Detection": 132,
        "Functional_Correspondence": 130,
        "IQ_Test": 150,
        "Jigsaw": 150,
        "Multi-view_Reasoning": 133,
        "Object_Localization": 122,
        "Relative_Depth": 124,
        "Relative_Reflectance": 134,
        "Semantic_Correspondence": 139,
        "Spatial_Relation": 143,
        "Visual_Correspondence": 172,
        "Visual_Similarity": 135,
    }
    if len(records) != 1901 or Counter(row["config"] for row in records) != Counter(expected_counts):
        raise RuntimeError("BLINK val count contract failed")
    _write_manifest(
        output,
        records,
        {
            "dataset": "BLINK-Benchmark/BLINK",
            "revision": REVISIONS["blink"],
            "split": "val",
            "scope": "full official val; public test labels are hidden",
        },
    )


def _vstar_choices(text: str) -> list[str]:
    matches = re.findall(r"(?m)^\s*\(([A-Z])\)\s*(.+?)\s*$", text)
    if not matches:
        matches = re.findall(r"(?m)^\s*([A-Z])[\.:]\s*(.+?)\s*$", text)
    letters = [letter for letter, _ in matches]
    expected_letters = [chr(ord("A") + index) for index in range(len(letters))]
    if letters != expected_letters:
        raise ValueError(f"Cannot parse contiguous VStar choices from: {text!r}")
    return [choice for _, choice in matches]


def prepare_vstar(raw: Path, output: Path) -> None:
    questions = raw / "test_questions.jsonl"
    if not questions.is_file():
        raise FileNotFoundError(questions)
    records: list[dict] = []
    for source_row_index, line in enumerate(questions.read_text(encoding="utf-8").splitlines()):
        if not line.strip():
            continue
        row = json.loads(line)
        image = (raw / str(row["image"])).resolve()
        if not image.is_file() or raw.resolve() not in image.parents:
            raise FileNotFoundError(image)
        with Image.open(image) as opened:
            width, height = opened.size
            opened.verify()
        choices = _vstar_choices(str(row["text"]))
        _expected(row["label"], choices)
        records.append(
            {
                "sample_id": str(row["question_id"]),
                "benchmark": "vstar",
                "split": "test",
                "config": str(row["category"]),
                "task": str(row["category"]),
                "row_index": len(records),
                "source_row_index": source_row_index,
                "images": [str(image)],
                "image_sizes": [[width, height]],
                "question": str(row["text"]).split("\n", 1)[0],
                "choices": choices,
                "answer": str(row["label"]),
                "prompt": str(row["text"]),
            }
        )
    if len(records) != 191:
        raise RuntimeError(f"Expected 191 VStar rows, got {len(records)}")
    _write_manifest(
        output,
        records,
        {
            "dataset": "craigwu/vstar_bench",
            "revision": REVISIONS["vstar"],
            "split": "test",
            "scope": "full official test",
        },
    )


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--raw-root", type=Path, default=Path("data/benchmarks/raw"))
    parser.add_argument("--output-root", type=Path, default=Path("data/benchmarks/prepared"))
    args = parser.parse_args()
    raw = args.raw_root.resolve()
    output = args.output_root.resolve()
    prepare_cvbench(raw / "CV-Bench", output / "cvbench_full")
    prepare_blink(raw / "BLINK", output / "blink_val")
    prepare_vstar(raw / "vstar_bench", output / "vstar_test")
    print(f"Prepared all benchmarks under {output}")


if __name__ == "__main__":
    main()