huggingface / huggingface/datasets
`load_dataset(...,save_infos=True)` not working without loading script
- Dominant language
- Python
- Stars
- 22k
- Forks
- 3.4k
- Avg merge
- 5d 7h
- Merged PRs (30d)
- 17
Description
### Describe the bug
It seems that saving a dataset infos back into the card file is not working for datasets without a loading script.
After tracking the problem a bit it looks like saving the infos uses `Builder.get_imported_module_dir()` as its destination directory.
Internally this is a call to `inspect.getfile()` but since the actual builder class used is dynamically created (cf. `datasets.load.configure_builder_class`) this method actually return te path to the parent builder class (e.g. `datasets.packaged_modules.json.JSON`).
### Steps to reproduce the bug
1. Have a local dataset without any loading script
2. Make sure there are no dataset infos in the README.md
3. Load with `save_infos=True`
4. No change in the dataset README.md
5. A new README.md file is created in the directory of the parent builder class (e.g. for json in `.../site-packages/datasets/packaged_modules/json/README.md`)
### Expected behavior
The dataset README.md should be updated and no file should be created in the python environment.
### Environment info
- `datasets` version: 2.15.0
- Platform: Linux-6.2.0-37-generic-x86_64-with-glibc2.35
- Python version: 3.10.12
- `huggingface_hub` version: 0.19.4
- PyArrow version: 14.0.1
- Pandas version: 2.1.3
- `fsspec` version: 2023.6.0
Contributor guide
Assessment
This issue has not been assessed yet.