`cache_dir` option in `download_config` in `load_dataset` is not respected

[TsXor]

Describe the bug

Downloaded files still go to ~/.cache/huggingface/hub/ even if I specified cache_dir option in download_config in load_dataset.

Steps to reproduce the bug

Run my freshly written script and found that downloaded files did not go where I want.

'''
下载OpenWebText数据集，允许使用代理
'''

if __name__ == '__main__':
    import argparse
    parser = argparse.ArgumentParser(description='Download TikToken Files')
    parser.add_argument('--output-path', required=True, metavar='PATH', help='输出目录')
    parser.add_argument('--mirror', required=False, metavar='URL', help='HF镜像网址，例如：https://hf-mirror.com')
    parser.add_argument('--proxy', required=False, metavar='URL', help='代理网址')
    args = parser.parse_args()
else: args = None


import os
import shutil
from pathlib import Path
from typing import cast


if __name__ == '__main__':
    assert args is not None
    output_path = Path(args.output_path).resolve()
    proxy_url = None if args.proxy is None else str(args.proxy)
    mirror_url = None if args.mirror is None else str(args.mirror)

    output_path.mkdir(parents=True, exist_ok=True)
    download_cache_dir = output_path / 'download_cache'
    read_cache_dir = output_path / 'read_cache'
    save_dir = output_path / 'saved'
    complete_mark = output_path / 'completed'

    def clear_cache():
        shutil.rmtree(download_cache_dir)
        shutil.rmtree(read_cache_dir)

    def download_and_save():
        if mirror_url is not None:
            os.environ["HF_ENDPOINT"] = mirror_url

        from datasets import DownloadConfig, load_dataset

        if proxy_url is not None: proxy_dict = { "http": proxy_url, "https": proxy_url }
        else: proxy_dict = None

        dataset = load_dataset(
            'openwebtext',
            cache_dir=str(read_cache_dir),
            download_config=DownloadConfig(cache_dir=download_cache_dir, proxies=proxy_dict)
        )
        dataset.save_to_disk(save_dir)

    if complete_mark.is_file():
        print('OpenWebText is already downloaded')
        clear_cache()
    else:
        download_and_save()
        complete_mark.touch(exist_ok=True)
        clear_cache()

Expected behavior

Downloaded files goes to where I specified in download_config.

Environment info

> uv run datasets-cli env

Copy-and-paste the text below in your GitHub issue.

- `datasets` version: 4.6.0
- Platform: Windows-11-10.0.26200-SP0
- Python version: 3.14.3
- `huggingface_hub` version: 1.5.0
- PyArrow version: 23.0.1
- Pandas version: 3.0.1
- `fsspec` version: 2026.2.0

[TsXor]

Actually, I used debugger to step into the code of datasets, and found load_dataset calls into cached_path, where it calls into hf_hub_download method for hf:// links:

datasets/src/datasets/utils/file_utils.py

Lines 182 to 195 in 6e8eaee

	output_path = huggingface_hub.HfApi(
	endpoint=config.HF_ENDPOINT,
	token=download_config.token,
	library_name="datasets",
	library_version=__version__,
	user_agent=get_datasets_user_agent(download_config.user_agent),
	).hf_hub_download(
	repo_id=resolved_path.repo_id,
	repo_type=resolved_path.repo_type,
	revision=resolved_path.revision,
	filename=resolved_path.path_in_repo,
	force_download=download_config.force_download,
	proxies=download_config.proxies,
	)

Some breakpoint testing indicated that hf_hub_download did the download job, and function signature says that it accepts cache_dir and local_dir as argument (which I'm pretty sure that it will control where download goes).

However, as you can see, the function call in the code above did not use these arguments, and the download function hf_hub_download has no idea where the download should go so it uses the default path and damned it puts all downloaded data to my user directory and is filling up my poor C drive.

It should be quite easy to solve this issue, as download_config is present at call site, and we only need to add several lines of code. The only question left is whether such behavior is intentionally designed. If so, please close this issue and forgive my ignorance.

[lhoestq]

That's a good point, and a PR to fix this would be very welcome. In the meantime you can set the HF_HUB_CACHE env variable, which tells where the cache for hub files lives