MedRCube.py

import json
import os
from typing import Any, Dict, Iterator, Tuple

import datasets
from datasets import Features, Image, Value


_DESCRIPTION = """MedRCube: A Multidimensional Framework for Fine-Grained and In-Depth Evaluation of MLLMs in Medical Imaging.

This loader discovers all `test.json` files under the dataset directory and emits a single `test` split.

Notes:
- Each example is multiple-choice VQA with 4 options (A-D).
- For license-restricted sources, images are not redistributed; those examples have `restricted=True` and `image=None`.
"""


class MedRCube(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    def _info(self) -> datasets.DatasetInfo:
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=Features(
                {
                    "id": Value("string"),
                    "dataset": Value("string"),
                    "image": Image(),
                    "image_path": Value("string"),
                    "restricted": Value("bool"),
                    "original_task": Value("string"),
                    "task": Value("string"),
                    "question": Value("string"),
                    "option_A": Value("string"),
                    "option_B": Value("string"),
                    "option_C": Value("string"),
                    "option_D": Value("string"),
                    "gt_answer": Value("string"),
                    "correct_index": Value("int32"),
                    "parts": Value("string"),
                    "modality": Value("string"),
                    # kept as unstructured JSON-compatible object
                    "metadata": datasets.Value("string"),
                }
            ),
            supervised_keys=None,
        )

    def _split_generators(self, dl_manager: datasets.DownloadManager):
        # In an HF snapshot, the data files live in the same repo root.
        # This script should be placed at the repo root (or in a folder that still
        # shares the same root as the dataset directories).
        repo_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"data_dir": repo_root},
            )
        ]

    def _generate_examples(self, data_dir: str) -> Iterator[Tuple[int, Dict[str, Any]]]:
        idx = 0
        for root, _, files in os.walk(data_dir):
            if "test.json" not in files:
                continue
            json_path = os.path.join(root, "test.json")
            with open(json_path, "r", encoding="utf-8") as f:
                items = json.load(f)
            if not isinstance(items, list):
                continue

            for item in items:
                image_path = item.get("image_path") or ""
                restricted = bool(item.get("restricted", False))

                abs_image_path = ""
                if image_path:
                    abs_image_path = (
                        os.path.normpath(os.path.join(root, image_path))
                        if not os.path.isabs(image_path)
                        else image_path
                    )

                image = None
                if abs_image_path and (not restricted) and os.path.exists(abs_image_path):
                    image = abs_image_path

                ex = dict(item)
                ex["image_path"] = abs_image_path
                ex["restricted"] = restricted
                ex["image"] = image
                ex["metadata"] = json.dumps(ex.get("metadata", {}), ensure_ascii=False)

                yield idx, ex
                idx += 1