load_dataset.py

import io, json, tarfile
import datasets

_URLPATTERN = "data/{:05d}.tar"              # 改成你自己的路径模式

class MyConfig(datasets.BuilderConfig):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

class MyDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [MyConfig(name="default", version=datasets.Version("1.0.0"))]

    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features({
                "video": datasets.Video(),          # mp4 -> Video() 解码
                "json":  datasets.Value("string"),  # 整块 JSON 存成纯字符串
            }),
        )

    def _split_generators(self, dl_manager):
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"tar_paths": [_URLPATTERN.format(i) for i in range(100)]},
            ),
        ]

    def _generate_examples(self, tar_paths):
        idx = 0
        for tar_path in tar_paths:
            with tarfile.open(tar_path, "r") as tar:
                members = {m.name: tar.extractfile(m).read() for m in tar.getmembers()}
                # 这里假设 <key>.mp4 / <key>.json；可按实际规则改
                for key in set(p.split(".")[0] for p in members):
                    mp4_bytes  = io.BytesIO(members[f"{key}.mp4"])
                    json_str   = members[f"{key}.json"].decode("utf-8")   # ← 直接当字符串
                    yield idx, {"video": {"path": mp4_bytes, "bytes": mp4_bytes.getvalue()},
                                "json": json_str}
                    idx += 1