id
int64 599M
3.26B
| number
int64 1
7.7k
| title
stringlengths 1
290
| body
stringlengths 0
228k
⌀ | state
stringclasses 2
values | html_url
stringlengths 46
51
| created_at
timestamp[s]date 2020-04-14 10:18:02
2025-07-23 08:04:53
| updated_at
timestamp[s]date 2020-04-27 16:04:17
2025-07-23 18:53:44
| closed_at
timestamp[s]date 2020-04-14 12:01:40
2025-07-23 16:44:42
⌀ | user
dict | labels
listlengths 0
4
| is_pull_request
bool 2
classes | comments
listlengths 0
0
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|
2,965,347,685 | 7,494 |
Broken links in pdf loading documentation
|
### Describe the bug
Hi, just a couple of small issues I ran into while reading the docs for [loading pdf data](https://huggingface.co/docs/datasets/main/en/document_load):
1. The link for the [`Create a pdf dataset`](https://huggingface.co/docs/datasets/main/en/document_load#pdffolder) points to https://huggingface.co/docs/datasets/main/en/pdf_dataset instead of https://huggingface.co/docs/datasets/main/en/document_dataset and hence gives a 404 error.
2. At the top of the page, it's mentioned that to work with pdf datasets we need to have the `pdfplumber` package installed but the link to its installation guide points to `pytorch/vision` [installation instructions](https://github.com/pytorch/vision#installation) instead of `pdfplumber`'s [guide](https://github.com/jsvine/pdfplumber#installation)
I love the work on enabling pdf dataset support and these small tweaks would help everyone navigate the docs better. Thanks!
### Steps to reproduce the bug
The issue is on the [Load Document Data](https://huggingface.co/docs/datasets/main/en/document_load) page of the datasets docs.
### Expected behavior
1. For solving the first issue, I went through the [source .mdx code](https://github.com/huggingface/datasets/blob/main/docs/source/document_load.mdx?plain=1#L188) of the datasets docs and found that the link is pointing to `./pdf_dataset` instead of `./document_dataset`
2. For the second issue, I went through the [source .mdx code](https://github.com/huggingface/datasets/blob/main/docs/source/document_load.mdx?plain=1#L13) of the datasets docs and found that the link is `pytorch/vision` [installation instructions](https://github.com/pytorch/vision#installation) instead of `pdfplumber`'s [guide](https://github.com/jsvine/pdfplumber#installation)
Just replacing these two links should fix the bugs
### Environment info
datasets v3.5.0 (main at the time of writing)
|
closed
|
https://github.com/huggingface/datasets/issues/7494
| 2025-04-02T06:45:22 | 2025-04-15T13:36:25 | 2025-04-15T13:36:04 |
{
"login": "VyoJ",
"id": 75789232,
"type": "User"
}
|
[] | false |
[] |
2,964,025,179 | 7,493 |
push_to_hub does not upload videos
|
### Describe the bug
Hello,
I would like to upload a video dataset (some .mp4 files and some segments within them), i.e. rows correspond to subsequences from videos. Videos might be referenced by several rows.
I created a dataset locally and it references the videos and the video readers can read them correctly. I use push_to_hub() to upload the dataset to the hub.
Expectation: A user uses `load_dataset` and can load the videos.
However, the videos seem to be just referenced via paths on the computer and not uploaded to the hub. Therefore a target user cannot load the videos in the dataset.
### Steps to reproduce the bug
1. create a video dataset with paths e.g. { ["videos"]: [path1, path2, ...]}
2. dataset.push_to_hub
3. on a different computer (or same pc if relative paths are used in a different folder):
```
dataset = load_dataset("siplab/egosim", split="train")
video = dataset[0]["video_head"]
```
3. will fail
### Expected behavior
Expectation: A user uses `load_dataset` and can load the videos.
### Environment info
datasets 3.1.0
Python 3.8.18
|
open
|
https://github.com/huggingface/datasets/issues/7493
| 2025-04-01T17:00:20 | 2025-04-15T12:34:23 | null |
{
"login": "DominikVincent",
"id": 9339403,
"type": "User"
}
|
[] | false |
[] |
2,959,088,568 | 7,492 |
Closes #7457
|
This PR updates the documentation to include the HF_DATASETS_CACHE environment variable, which allows users to customize the cache location for datasets—similar to HF_HUB_CACHE for models.
|
closed
|
https://github.com/huggingface/datasets/pull/7492
| 2025-03-30T20:41:20 | 2025-04-13T22:05:07 | 2025-04-13T22:05:07 |
{
"login": "Harry-Yang0518",
"id": 129883215,
"type": "User"
}
|
[] | true |
[] |
2,959,085,647 | 7,491 |
docs: update cache.mdx to include HF_DATASETS_CACHE documentation
| null |
closed
|
https://github.com/huggingface/datasets/pull/7491
| 2025-03-30T20:35:03 | 2025-03-30T20:36:40 | 2025-03-30T20:36:40 |
{
"login": "Harry-Yang0518",
"id": 129883215,
"type": "User"
}
|
[] | true |
[] |
2,958,826,222 | 7,490 |
(refactor) remove redundant logic in _check_valid_index_key
|
This PR contributes a minor refactor, in a small function in `src/datasets/formatting/formatting.py`. No change in logic.
In the original code, there are separate if-conditionals for `isinstance(key, range)` and `isinstance(key, Iterable)`, with essentially the same logic.
This PR combines these two using a single if statement.
**Considerations**
1. Although range in python is guaranteed to have integers, internally calling `int()` on an object that is already an int is negligible. (In python it returns the original object. It doesn't create a new integer object or perform any actual conversion)
2. Technically a range is already an Iterable, and we could just do `isinstance(key, Iterable)` but I explicitly did `isinstance(key, (range, Iterable))` just to be super obvious and consistent that both cases are handled because I see `slice, range, Iterable` everywhere in this `formatting.py`
3. This PR removes the `if len(key)>0` conditional. I think it is cleaner to have it this way for three reasons.
- There was originally no else statement and the code would have failed silently anyway.
- The if len(key)>0 should be caught much earlier, rather than in `formatting.py`.
- There are actually multiple cases where this would fail, if len(key)>0, if key is non numeric or float, or if key is a list of lists. It's clunky to state all this and the error be thrown during max or indexing.
**Previous PR and Issues Checks**
1. No known PR or Issues (both closed or open) in hf datasets repository
**Tests**
1. Tested using Dataset (load_dataset("wikitext", "wikitext-103-raw-v1")), Pytorch DataLoader, with a Pytorch BatchSampler (list of indexes returned instead of single index).
|
open
|
https://github.com/huggingface/datasets/pull/7490
| 2025-03-30T11:45:42 | 2025-03-30T11:50:22 | null |
{
"login": "suzyahyah",
"id": 2980993,
"type": "User"
}
|
[] | true |
[] |
2,958,204,763 | 7,489 |
fix: loading of datasets from Disk(#7373)
|
Fixes dataset loading from disk by ensuring that memory maps and streams are properly closed.
For more details, see https://github.com/huggingface/datasets/issues/7373.
|
open
|
https://github.com/huggingface/datasets/pull/7489
| 2025-03-29T16:22:58 | 2025-04-24T16:36:36 | null |
{
"login": "sam-hey",
"id": 40773225,
"type": "User"
}
|
[] | true |
[] |
2,956,559,358 | 7,488 |
Support underscore int read instruction
|
close https://github.com/huggingface/datasets/issues/7481
|
closed
|
https://github.com/huggingface/datasets/pull/7488
| 2025-03-28T16:01:15 | 2025-03-28T16:20:44 | 2025-03-28T16:20:43 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,956,533,448 | 7,487 |
Write pdf in map
|
Fix this error when mapping a PDF dataset
```
pyarrow.lib.ArrowInvalid: Could not convert <pdfplumber.pdf.PDF object at 0x13498ee40> with type PDF: did not recognize Python value type when inferring an Arrow data type
```
and also let map() outputs be lists of images or pdfs
|
closed
|
https://github.com/huggingface/datasets/pull/7487
| 2025-03-28T15:49:25 | 2025-03-28T17:09:53 | 2025-03-28T17:09:51 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,954,042,179 | 7,486 |
`shared_datadir` fixture is missing
|
### Describe the bug
Running the tests for the latest release fails due to missing `shared_datadir` fixture.
### Steps to reproduce the bug
Running `pytest` while building a package for Arch Linux leads to these errors:
```
==================================== ERRORS ====================================
_________ ERROR at setup of test_pdf_feature_encode_example[<lambda>1] _________
[gw44] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 8
@require_pdfplumber
@pytest.mark.parametrize(
"build_example",
[
lambda pdf_path: pdf_path,
lambda pdf_path: open(pdf_path, "rb").read(),
lambda pdf_path: {"path": pdf_path},
lambda pdf_path: {"path": pdf_path, "bytes": None},
lambda pdf_path: {"path": pdf_path, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"path": None, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"bytes": open(pdf_path, "rb").read()},
],
)
def test_pdf_feature_encode_example(shared_datadir, build_example):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:8
_________ ERROR at setup of test_pdf_feature_encode_example[<lambda>2] _________
[gw44] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 8
@require_pdfplumber
@pytest.mark.parametrize(
"build_example",
[
lambda pdf_path: pdf_path,
lambda pdf_path: open(pdf_path, "rb").read(),
lambda pdf_path: {"path": pdf_path},
lambda pdf_path: {"path": pdf_path, "bytes": None},
lambda pdf_path: {"path": pdf_path, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"path": None, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"bytes": open(pdf_path, "rb").read()},
],
)
def test_pdf_feature_encode_example(shared_datadir, build_example):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:8
_________ ERROR at setup of test_pdf_feature_encode_example[<lambda>3] _________
[gw44] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 8
@require_pdfplumber
@pytest.mark.parametrize(
"build_example",
[
lambda pdf_path: pdf_path,
lambda pdf_path: open(pdf_path, "rb").read(),
lambda pdf_path: {"path": pdf_path},
lambda pdf_path: {"path": pdf_path, "bytes": None},
lambda pdf_path: {"path": pdf_path, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"path": None, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"bytes": open(pdf_path, "rb").read()},
],
)
def test_pdf_feature_encode_example(shared_datadir, build_example):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:8
_________ ERROR at setup of test_pdf_feature_encode_example[<lambda>4] _________
[gw44] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 8
@require_pdfplumber
@pytest.mark.parametrize(
"build_example",
[
lambda pdf_path: pdf_path,
lambda pdf_path: open(pdf_path, "rb").read(),
lambda pdf_path: {"path": pdf_path},
lambda pdf_path: {"path": pdf_path, "bytes": None},
lambda pdf_path: {"path": pdf_path, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"path": None, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"bytes": open(pdf_path, "rb").read()},
],
)
def test_pdf_feature_encode_example(shared_datadir, build_example):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:8
_________ ERROR at setup of test_pdf_feature_encode_example[<lambda>5] _________
[gw44] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 8
@require_pdfplumber
@pytest.mark.parametrize(
"build_example",
[
lambda pdf_path: pdf_path,
lambda pdf_path: open(pdf_path, "rb").read(),
lambda pdf_path: {"path": pdf_path},
lambda pdf_path: {"path": pdf_path, "bytes": None},
lambda pdf_path: {"path": pdf_path, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"path": None, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"bytes": open(pdf_path, "rb").read()},
],
)
def test_pdf_feature_encode_example(shared_datadir, build_example):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:8
_________ ERROR at setup of test_pdf_feature_encode_example[<lambda>6] _________
[gw44] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 8
@require_pdfplumber
@pytest.mark.parametrize(
"build_example",
[
lambda pdf_path: pdf_path,
lambda pdf_path: open(pdf_path, "rb").read(),
lambda pdf_path: {"path": pdf_path},
lambda pdf_path: {"path": pdf_path, "bytes": None},
lambda pdf_path: {"path": pdf_path, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"path": None, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"bytes": open(pdf_path, "rb").read()},
],
)
def test_pdf_feature_encode_example(shared_datadir, build_example):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:8
_______________ ERROR at setup of test_dataset_with_pdf_feature ________________
[gw44] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 34
@require_pdfplumber
def test_dataset_with_pdf_feature(shared_datadir):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:34
_________ ERROR at setup of test_pdf_feature_encode_example[<lambda>0] _________
[gw46] linux -- Python 3.13.2 /build/python-datasets/src/datasets-3.5.0/test-env/bin/python
file /build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py, line 8
@require_pdfplumber
@pytest.mark.parametrize(
"build_example",
[
lambda pdf_path: pdf_path,
lambda pdf_path: open(pdf_path, "rb").read(),
lambda pdf_path: {"path": pdf_path},
lambda pdf_path: {"path": pdf_path, "bytes": None},
lambda pdf_path: {"path": pdf_path, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"path": None, "bytes": open(pdf_path, "rb").read()},
lambda pdf_path: {"bytes": open(pdf_path, "rb").read()},
],
)
def test_pdf_feature_encode_example(shared_datadir, build_example):
E fixture 'shared_datadir' not found
> available fixtures: _hf_gated_dataset_repo_txt_data, arrow_file, arrow_path, audio_file, bz2_csv_path, bz2_file, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, ci_hfh_hf_hub_url, ci_hub_config, cleanup_repo, csv2_path, csv_path, data_dir_with_hidden_files, dataset, dataset_dict, disable_implicit_token, disable_tqdm_output, doctest_namespace, geoparquet_path, gz_file, hf_api, hf_gated_dataset_repo_txt_data, hf_private_dataset_repo_txt_data, hf_private_dataset_repo_txt_data_, hf_private_dataset_repo_zipped_img_data, hf_private_dataset_repo_zipped_img_data_, hf_private_dataset_repo_zipped_txt_data, hf_private_dataset_repo_zipped_txt_data_, hf_token, image_file, json_dict_of_lists_path, json_list_of_dicts_path, jsonl2_path, jsonl_312_path, jsonl_gz_path, jsonl_path, jsonl_str_path, lz4_file, mock_fsspec, mockfs, monkeypatch, parquet_path, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, set_ci_hub_access_token, set_sqlalchemy_silence_uber_warning, set_test_cache_config, set_update_download_counts_to_false, seven_zip_file, sqlite_path, tar_file, tar_jsonl_path, tar_nested_jsonl_path, temporary_repo, tensor_file, testrun_uid, text2_path, text_dir, text_dir_with_unsupported_extension, text_file, text_file_content, text_gz_path, text_path, text_path_with_unicode_new_lines, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tmpfs, worker_id, xml_file, xz_file, zero_time_out_for_remote_code, zip_csv_path, zip_csv_with_dir_path, zip_file, zip_image_path, zip_jsonl_path, zip_jsonl_with_dir_path, zip_nested_jsonl_path, zip_text_path, zip_text_with_dir_path, zip_unsupported_ext_path, zip_uppercase_csv_path, zstd_file
> use 'pytest --fixtures [testpath]' for help on them.
/build/python-datasets/src/datasets-3.5.0/tests/features/test_pdf.py:8
```
### Expected behavior
All fixtures used in tests should be available.
### Environment info
Arch Linux build system, building the [python-datasets](https://gitlab.archlinux.org/archlinux/packaging/packages/python-datasets) package.
There are actually [many deselected tests](https://gitlab.archlinux.org/archlinux/packaging/packages/python-datasets/-/blob/6f97957f0c326cc7b3da6b7f12326305bcaef374/PKGBUILD#L66-148) which were failing on previous releases, but these errors popped up in 3.5.0.
|
closed
|
https://github.com/huggingface/datasets/issues/7486
| 2025-03-27T18:17:12 | 2025-03-27T19:49:11 | 2025-03-27T19:49:10 |
{
"login": "lahwaacz",
"id": 1289205,
"type": "User"
}
|
[] | false |
[] |
2,953,696,519 | 7,485 |
set dev version
| null |
closed
|
https://github.com/huggingface/datasets/pull/7485
| 2025-03-27T16:39:34 | 2025-03-27T16:41:59 | 2025-03-27T16:39:42 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,953,677,168 | 7,484 |
release: 3.5.0
| null |
closed
|
https://github.com/huggingface/datasets/pull/7484
| 2025-03-27T16:33:27 | 2025-03-27T16:35:44 | 2025-03-27T16:34:22 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,951,856,468 | 7,483 |
Support skip_trying_type
|
This PR addresses Issue #7472
cc: @lhoestq
|
closed
|
https://github.com/huggingface/datasets/pull/7483
| 2025-03-27T07:07:20 | 2025-04-29T04:14:57 | 2025-04-09T09:53:10 |
{
"login": "yoshitomo-matsubara",
"id": 11156001,
"type": "User"
}
|
[] | true |
[] |
2,950,890,368 | 7,482 |
Implement capability to restore non-nullability in Features
|
This PR attempts to keep track of non_nullable pyarrow fields when converting a `pa.Schema` to `Features`. At the same time, when outputting the `arrow_schema`, the original non-nullable fields are restored. This allows for more consistent behavior and avoids breaking behavior as illustrated in #7479.
I am by no means a pyarrow expert so some logic in `find_non_nullable_fields` may not perfect. Not sure if more logic (type checks) are needed for deep-checking a given schema. Maybe there are other pyarrow structures that need to be covered?
Tests are added, but again, these may not have sufficient coverage in terms of pyarrow structure types.
closes #7479
|
closed
|
https://github.com/huggingface/datasets/pull/7482
| 2025-03-26T22:16:09 | 2025-05-15T15:00:59 | 2025-05-15T15:00:59 |
{
"login": "BramVanroy",
"id": 2779410,
"type": "User"
}
|
[] | true |
[] |
2,950,692,971 | 7,481 |
deal with python `10_000` legal number in slice syntax
|
### Feature request
```
In [6]: ds = datasets.load_dataset("HuggingFaceH4/ultrachat_200k", split="train_sft[:1000]")
In [7]: ds = datasets.load_dataset("HuggingFaceH4/ultrachat_200k", split="train_sft[:1_000]")
[dozens of frames skipped]
File /usr/local/lib/python3.10/dist-packages/datasets/arrow_reader.py:444, in _str_to_read_instruction(spec)
442 res = _SUB_SPEC_RE.match(spec)
443 if not res:
--> 444 raise ValueError(f"Unrecognized instruction format: {spec}")
ValueError: Unrecognized instruction format: train_sft[:1_000]
```
It took me a while to understand what the problem was. But apparently `pyarrow` doesn't allow python numbers that may include `_` as in `1_000`. The `_` aids readability since `10_000_000` vs `10000000` is obviously easier to grasp of what the actual number is.
Feature request:
ideally `datasets` being a python module will do the right thing and convert python numbers into whatever pyarrow supports - in this case stripping `_`s.
Second best it'd err and tell the user that using numbers with `_` in split slices is not acceptible, so that the user won't have to deal with a huge pyarrow assert they know nothing about.
Thank you!
|
closed
|
https://github.com/huggingface/datasets/issues/7481
| 2025-03-26T20:10:54 | 2025-03-28T16:20:44 | 2025-03-28T16:20:44 |
{
"login": "sfc-gh-sbekman",
"id": 196988264,
"type": "User"
}
|
[
{
"name": "enhancement",
"color": "a2eeef"
}
] | false |
[] |
2,950,315,214 | 7,480 |
HF_DATASETS_CACHE ignored?
|
### Describe the bug
I'm struggling to get things to respect HF_DATASETS_CACHE.
Rationale: I'm on a system that uses NFS for homedir, so downloading to NFS is expensive, slow, and wastes valuable quota compared to local disk. Instead, it seems to rely mostly on HF_HUB_CACHE.
Current version: 3.2.1dev. In the process of testing 3.4.0
### Steps to reproduce the bug
[Currently writing using datasets 3.2.1dev. Will follow up with 3.4.0 results]
dump.py:
```python
from datasets import load_dataset
dataset = load_dataset("HuggingFaceFW/fineweb", name="sample-100BT", split="train")
```
Repro steps
```bash
# ensure no cache
$ mv ~/.cache/huggingface ~/.cache/huggingface.bak
$ export HF_DATASETS_CACHE=/tmp/roller/datasets
$ rm -rf ${HF_DATASETS_CACHE}
$ env | grep HF | grep -v TOKEN
HF_DATASETS_CACHE=/tmp/roller/datasets
$ python dump.py
# (omitted for brevity)
# (while downloading)
$ du -hcs ~/.cache/huggingface/hub
18G hub
18G total
# (after downloading)
$ du -hcs ~/.cache/huggingface/hub
```
It's a shame because datasets supports s3 (which I could really use right now) but hub does not.
### Expected behavior
* ~/.cache/huggingface/hub stays empty
* /tmp/roller/datasets becomes full of stuff
### Environment info
[Currently writing using datasets 3.2.1dev. Will follow up with 3.4.0 results]
|
open
|
https://github.com/huggingface/datasets/issues/7480
| 2025-03-26T17:19:34 | 2025-04-28T10:16:16 | null |
{
"login": "stephenroller",
"id": 31896,
"type": "User"
}
|
[] | false |
[] |
2,950,235,396 | 7,479 |
Features.from_arrow_schema is destructive
|
### Describe the bug
I came across this, perhaps niche, bug where `Features` does not/cannot account for pyarrow's `nullable=False` option in Fields. Interestingly, I found that in regular "flat" fields this does not necessarily lead to conflicts, but when a non-nullable field is in a struct, an incompatibility arises.
It's not easy to explain in words, so the minimal example below should help I hope.
Note that I suggest a solution in the comments in the code, simply allowing `Dataset.to_parquet` to allow for a `schema` argument which, when provided, will override the default ds.features.arrow_schema.
### Steps to reproduce the bug
```python
import os
from datasets import Dataset, Features
import pyarrow as pa
import pyarrow.parquet as pq
# HF datasets is destructive when you call Features.from_arrow_schema(schema) on a schema
# because it will not account for nullable and non-nullable fields in structs (it will always allow nullable)
# Reloading the same dataset with the original schema will raise an error because the schema is not the same anymore
non_nullable_schema = pa.schema(
[
pa.field("text", pa.string(), nullable=False),
pa.field("meta",
pa.struct(
[
pa.field("date", pa.list_(pa.string()), nullable=False),
],
),
),
]
)
print("ORIGINAL SCHEMA")
print(non_nullable_schema)
print()
feats = Features.from_arrow_schema(non_nullable_schema)
print("FEATUR-IZED SCHEMA (nullable-restrictions are gone)")
print(feats.arrow_schema)
print()
ds = Dataset.from_dict(
{
"text": ["a", "b", "c"],
"meta": [{"date": ["2021-01-01"]}, {"date": ["2021-01-02"]}, {"date": ["2021-01-03"]}],
},
features=feats,
)
fname = "tmp.parquet"
# This is not possible: TypeError: pyarrow.parquet.core.ParquetWriter() got multiple values for keyword argument 'schema'
# Though I believe this would be the easiest fix: allow schema to be passed to to_parquet and overwrite the schema in the dataset
# ds.to_parquet(fname, schema=non_nullable_schema)
ds.to_parquet(fname)
try:
_ = pq.read_table(fname, schema=non_nullable_schema)
finally:
os.unlink(fname)
```
### Expected behavior
- Non-destructive behavior when converting an arrow schema to Features; or
- the ability to override the default arrow schema with a custom one
### Environment info
- `datasets` version: 3.2.0
- Platform: Linux-5.14.0-427.20.1.el9_4.x86_64-x86_64-with-glibc2.34
- Python version: 3.11.10
- `huggingface_hub` version: 0.27.1
- PyArrow version: 18.1.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.9.0
|
open
|
https://github.com/huggingface/datasets/issues/7479
| 2025-03-26T16:46:43 | 2025-03-26T16:46:58 | null |
{
"login": "BramVanroy",
"id": 2779410,
"type": "User"
}
|
[] | false |
[] |
2,948,993,461 | 7,478 |
update fsspec 2025.3.0
|
It appears there have been two releases of fsspec since this dependency was last updated, it would be great if Datasets could be updated so that it didn't hold back the usage of newer fsspec versions in consuming projects.
PR based on https://github.com/huggingface/datasets/pull/7352
|
closed
|
https://github.com/huggingface/datasets/pull/7478
| 2025-03-26T09:53:05 | 2025-03-28T19:15:54 | 2025-03-28T15:51:55 |
{
"login": "peteski22",
"id": 487783,
"type": "User"
}
|
[] | true |
[] |
2,947,169,460 | 7,477 |
What is the canonical way to compress a Dataset?
|
Given that Arrow is the preferred backend for a Dataset, what is a user supposed to do if they want concurrent reads, concurrent writes AND on-disk compression for a larger dataset?
Parquet would be the obvious answer except that there is no native support for writing sharded, parquet datasets concurrently [[1](https://github.com/huggingface/datasets/issues/7047)].
Am I missing something?
And if so, why is this not the standard/default way that `Dataset`'s work as they do in Xarray, Ray Data, Composer, etc.?
|
open
|
https://github.com/huggingface/datasets/issues/7477
| 2025-03-25T16:47:51 | 2025-04-03T09:13:11 | null |
{
"login": "eric-czech",
"id": 6130352,
"type": "User"
}
|
[] | false |
[] |
2,946,997,924 | 7,476 |
Priotitize json
|
`datasets` should load the JSON data in https://huggingface.co/datasets/facebook/natural_reasoning, not the PDF
|
closed
|
https://github.com/huggingface/datasets/pull/7476
| 2025-03-25T15:44:31 | 2025-03-25T15:47:00 | 2025-03-25T15:45:00 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,946,640,570 | 7,475 |
IterableDataset's state_dict shard_example_idx is always equal to the number of samples in a shard
|
### Describe the bug
I've noticed a strange behaviour with Iterable state_dict: the value of shard_example_idx is always equal to the amount of samples in a shard.
### Steps to reproduce the bug
I am reusing the example from the doc
```python
from datasets import Dataset
ds = Dataset.from_dict({"a": range(6)}).to_iterable_dataset(num_shards=1)
state_dict = None
# Iterate through the dataset and print examples
for idx, example in enumerate(ds):
print(example)
if idx == 2:
state_dict = ds.state_dict()
print("checkpoint")
break
print(state_dict)
```
Returns:
```
{'a': 0}
{'a': 1}
checkpoint
{'examples_iterable': {'shard_idx': 0, 'shard_example_idx': 6, 'type': 'ArrowExamplesIterable'}, 'epoch': 0}
```
### Expected behavior
shard_example_idx should be 2 instead of 6
If we run with num_shards=2, then shard_example_idx is 3 instead of 2 and so on.
### Environment info
- `datasets` version: 3.4.1
- Platform: macOS-14.6.1-arm64-arm-64bit
- Python version: 3.12.9
- `huggingface_hub` version: 0.29.3
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
|
closed
|
https://github.com/huggingface/datasets/issues/7475
| 2025-03-25T13:58:07 | 2025-05-06T14:22:19 | 2025-05-06T14:05:07 |
{
"login": "bruno-hays",
"id": 48770768,
"type": "User"
}
|
[] | false |
[] |
2,945,066,258 | 7,474 |
Remove conditions for Python < 3.9
|
This PR remove conditions for Python < 3.9.
|
closed
|
https://github.com/huggingface/datasets/pull/7474
| 2025-03-25T03:08:04 | 2025-04-16T00:11:06 | 2025-04-15T16:07:55 |
{
"login": "cyyever",
"id": 17618148,
"type": "User"
}
|
[] | true |
[] |
2,939,034,643 | 7,473 |
Webdataset data format problem
|
### Describe the bug
Please see https://huggingface.co/datasets/ejschwartz/idioms/discussions/1
Error code: FileFormatMismatchBetweenSplitsError
All three splits, train, test, and validation, use webdataset. But only the train split has more than one file. How can I force the other two splits to also be interpreted as being the webdataset format? (I don't think there is currently a way, but happy to be told that I am wrong.)
### Steps to reproduce the bug
```
import datasets
datasets.load_dataset("ejschwartz/idioms")
### Expected behavior
The dataset loads. Alternatively, there is a YAML syntax for manually specifying the format.
### Environment info
- `datasets` version: 3.2.0
- Platform: Linux-6.8.0-52-generic-x86_64-with-glibc2.35
- Python version: 3.10.12
- `huggingface_hub` version: 0.28.1
- PyArrow version: 19.0.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.9.0
|
closed
|
https://github.com/huggingface/datasets/issues/7473
| 2025-03-21T17:23:52 | 2025-03-21T19:19:58 | 2025-03-21T19:19:58 |
{
"login": "edmcman",
"id": 1017189,
"type": "User"
}
|
[] | false |
[] |
2,937,607,272 | 7,472 |
Label casting during `map` process is canceled after the `map` process
|
### Describe the bug
When preprocessing a multi-label dataset, I introduced a step to convert int labels to float labels as [BCEWithLogitsLoss](https://pytorch.org/docs/stable/generated/torch.nn.BCEWithLogitsLoss.html) expects float labels and forward function of models in transformers package internally use `BCEWithLogitsLoss`
However, the casting was canceled after `.map` process and the label values still use int values, which leads to an error
```
File "/home/yoshitomo/anaconda3/envs/torchdistill/lib/python3.10/site-packages/transformers/models/bert/modeling_bert.py", line 1711, in forward
loss = loss_fct(logits, labels)
File "/home/yoshitomo/anaconda3/envs/torchdistill/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/home/yoshitomo/anaconda3/envs/torchdistill/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
File "/home/yoshitomo/anaconda3/envs/torchdistill/lib/python3.10/site-packages/torch/nn/modules/loss.py", line 819, in forward
return F.binary_cross_entropy_with_logits(
File "/home/yoshitomo/anaconda3/envs/torchdistill/lib/python3.10/site-packages/torch/nn/functional.py", line 3628, in binary_cross_entropy_with_logits
return torch.binary_cross_entropy_with_logits(
RuntimeError: result type Float can't be cast to the desired output type Long
```
This seems like happening only when the original labels are int values (see examples below)
### Steps to reproduce the bug
If the original dataset uses a list of int labels, it will cancel the int->float casting
```python
from datasets import Dataset
data = {
'text': ['text1', 'text2', 'text3', 'text4'],
'labels': [[0, 1, 2], [3], [3, 4], [3]]
}
dataset = Dataset.from_dict(data)
label_set = set([label for labels in data['labels'] for label in labels])
label2idx = {label: idx for idx, label in enumerate(sorted(label_set))}
def multi_labels_to_ids(labels):
ids = [0.0] * len(label2idx)
for label in labels:
ids[label2idx[label]] = 1.0
return ids
def preprocess(examples):
result = {'sentence': [[0, 3, 4] for _ in range(len(examples['labels']))]}
print('"labels" are int', examples['labels'])
result['labels'] = [multi_labels_to_ids(l) for l in examples['labels']]
print('"labels" were converted to multi-label format with float values', result['labels'])
return result
preprocessed_dataset = dataset.map(preprocess, batched=True, remove_columns=['labels', 'text'])
print(preprocessed_dataset[0]['labels'])
# Output: "[1, 1, 1, 0, 0]"
# Expected: "[1.0, 1.0, 1.0, 0.0, 0.0]"
```
If the original dataset uses non-int labels, it works as expected.
```python
from datasets import Dataset
data = {
'text': ['text1', 'text2', 'text3', 'text4'],
'labels': [['label1', 'label2', 'label3'], ['label4'], ['label4', 'label5'], ['label4']]
}
dataset = Dataset.from_dict(data)
label_set = set([label for labels in data['labels'] for label in labels])
label2idx = {label: idx for idx, label in enumerate(sorted(label_set))}
def multi_labels_to_ids(labels):
ids = [0.0] * len(label2idx)
for label in labels:
ids[label2idx[label]] = 1.0
return ids
def preprocess(examples):
result = {'sentence': [[0, 3, 4] for _ in range(len(examples['labels']))]}
print('"labels" are int', examples['labels'])
result['labels'] = [multi_labels_to_ids(l) for l in examples['labels']]
print('"labels" were converted to multi-label format with float values', result['labels'])
return result
preprocessed_dataset = dataset.map(preprocess, batched=True, remove_columns=['labels', 'text'])
print(preprocessed_dataset[0]['labels'])
# Output: "[1.0, 1.0, 1.0, 0.0, 0.0]"
# Expected: "[1.0, 1.0, 1.0, 0.0, 0.0]"
```
Note that the only difference between these two examples is
> 'labels': [[0, 1, 2], [3], [3, 4], [3]]
v.s
> 'labels': [['label1', 'label2', 'label3'], ['label4'], ['label4', 'label5'], ['label4']]
### Expected behavior
Even if the original dataset uses a list of int labels, the int->float casting during `.map` process should not be canceled as shown in the above example
### Environment info
OS Ubuntu 22.04 LTS
Python 3.10.11
datasets v3.4.1
|
closed
|
https://github.com/huggingface/datasets/issues/7472
| 2025-03-21T07:56:22 | 2025-04-10T05:11:15 | 2025-04-10T05:11:14 |
{
"login": "yoshitomo-matsubara",
"id": 11156001,
"type": "User"
}
|
[] | false |
[] |
2,937,530,069 | 7,471 |
Adding argument to `_get_data_files_patterns`
|
### Feature request
How about adding if the user already know about the pattern?
https://github.com/huggingface/datasets/blob/a256b85cbc67aa3f0e75d32d6586afc507cf535b/src/datasets/data_files.py#L252
### Motivation
While using this load_dataset people might use 10M of images for the local files.
However, due to searching all the appropriate file pattern in fsspec, purely searching this pattern takes more than 10 hours (real use-case).
### Your contribution
Yeah I can make this happen if this seems valid. @lhoestq WDYT?
such like
```
def _get_data_files_patterns(pattern_resolver: Callable[[str], list[str]], patterns: PATTERNS) -> dict[str, list[str]]:
```
|
closed
|
https://github.com/huggingface/datasets/issues/7471
| 2025-03-21T07:17:53 | 2025-03-27T12:30:52 | 2025-03-26T07:26:27 |
{
"login": "SangbumChoi",
"id": 34004152,
"type": "User"
}
|
[
{
"name": "enhancement",
"color": "a2eeef"
}
] | false |
[] |
2,937,236,323 | 7,470 |
Is it possible to shard a single-sharded IterableDataset?
|
I thought https://github.com/huggingface/datasets/pull/7252 might be applicable but looking at it maybe not.
Say we have a process, eg. a database query, that can return data in slightly different order each time. So, the initial query needs to be run by a single thread (not to mention running multiple times incurs more cost too). But the results are also big enough that we don't want to materialize it entirely and instead stream it with an IterableDataset.
But after we have the results we want to split it up across workers to parallelize processing.
Is something like this possible to do?
Here's a failed attempt. The end result should be that each of the shards has unique data, but unfortunately with this attempt the generator gets run once in each shard and the results end up with duplicates...
```
import random
import datasets
def gen():
print('RUNNING GENERATOR!')
items = list(range(10))
random.shuffle(items)
yield from items
ds = datasets.IterableDataset.from_generator(gen)
print('dataset contents:')
for item in ds:
print(item)
print()
print('dataset contents (2):')
for item in ds:
print(item)
print()
num_shards = 3
def sharded(shard_id):
for i, example in enumerate(ds):
if i % num_shards in shard_id:
yield example
ds1 = datasets.IterableDataset.from_generator(
sharded, gen_kwargs={'shard_id': list(range(num_shards))}
)
for shard in range(num_shards):
print('shard', shard)
for item in ds1.shard(num_shards, shard):
print(item)
```
|
closed
|
https://github.com/huggingface/datasets/issues/7470
| 2025-03-21T04:33:37 | 2025-05-09T22:51:46 | 2025-03-26T06:49:28 |
{
"login": "jonathanasdf",
"id": 511073,
"type": "User"
}
|
[] | false |
[] |
2,936,606,080 | 7,469 |
Custom split name with the web interface
|
### Describe the bug
According the doc here: https://huggingface.co/docs/hub/datasets-file-names-and-splits#custom-split-name
it should infer the split name from the subdir of data or the beg of the name of the files in data.
When doing this manually through web upload it does not work. it uses "train" as a unique split.
example: https://huggingface.co/datasets/eole-nlp/estimator_chatml
### Steps to reproduce the bug
follow the link above
### Expected behavior
there should be two splits "mlqe" and "1720_da"
### Environment info
website
|
closed
|
https://github.com/huggingface/datasets/issues/7469
| 2025-03-20T20:45:59 | 2025-03-21T07:20:37 | 2025-03-21T07:20:37 |
{
"login": "vince62s",
"id": 15141326,
"type": "User"
}
|
[] | false |
[] |
2,934,094,103 | 7,468 |
function `load_dataset` can't solve folder path with regex characters like "[]"
|
### Describe the bug
When using the `load_dataset` function with a folder path containing regex special characters (such as "[]"), the issue occurs due to how the path is handled in the `resolve_pattern` function. This function passes the unprocessed path directly to `AbstractFileSystem.glob`, which supports regular expressions. As a result, the globbing mechanism interprets these characters as regex patterns, leading to a traversal of the entire disk partition instead of confining the search to the intended directory.
### Steps to reproduce the bug
just create a folder like `E:\[D_DATA]\koch_test`, then `load_dataset("parquet", data_dir="E:\[D_DATA]\\test", split="train")`
it will keep searching the whole disk.
I add two `print` in `glob` and `resolve_pattern` to see the path
### Expected behavior
it should load the dataset as in normal folders
### Environment info
- `datasets` version: 3.3.2
- Platform: Windows-10-10.0.22631-SP0
- Python version: 3.10.16
- `huggingface_hub` version: 0.29.1
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
|
open
|
https://github.com/huggingface/datasets/issues/7468
| 2025-03-20T05:21:59 | 2025-03-25T10:18:12 | null |
{
"login": "Hpeox",
"id": 89294013,
"type": "User"
}
|
[] | false |
[] |
2,930,067,107 | 7,467 |
load_dataset with streaming hangs on parquet datasets
|
### Describe the bug
When I try to load a dataset with parquet files (e.g. "bigcode/the-stack") the dataset loads, but python interpreter can't exit and hangs
### Steps to reproduce the bug
```python3
import datasets
print('Start')
dataset = datasets.load_dataset("bigcode/the-stack", data_dir="data/yaml", streaming=True, split="train")
it = iter(dataset)
next(it)
print('Finish')
```
The program prints finish but doesn't exit and hangs indefinitely.
I tried this on two different machines and several datasets.
### Expected behavior
The program exits successfully
### Environment info
datasets==3.4.1
Python 3.12.9.
MacOS and Ubuntu Linux
|
open
|
https://github.com/huggingface/datasets/issues/7467
| 2025-03-18T23:33:54 | 2025-03-25T10:28:04 | null |
{
"login": "The0nix",
"id": 10550252,
"type": "User"
}
|
[] | false |
[] |
2,928,661,327 | 7,466 |
Fix local pdf loading
|
fir this error when accessing a local pdf
```
File ~/.pyenv/versions/3.12.2/envs/hf-datasets/lib/python3.12/site-packages/pdfminer/psparser.py:220, in PSBaseParser.seek(self, pos)
218 """Seeks the parser to the given position."""
219 log.debug("seek: %r", pos)
--> 220 self.fp.seek(pos)
221 # reset the status for nextline()
222 self.bufpos = pos
ValueError: seek of closed file
```
|
closed
|
https://github.com/huggingface/datasets/pull/7466
| 2025-03-18T14:09:06 | 2025-03-18T14:11:52 | 2025-03-18T14:09:21 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,926,478,838 | 7,464 |
Minor fix for metadata files in extension counter
| null |
closed
|
https://github.com/huggingface/datasets/pull/7464
| 2025-03-17T21:57:11 | 2025-03-18T15:21:43 | 2025-03-18T15:21:41 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,925,924,452 | 7,463 |
Adds EXR format to store depth images in float32
|
This PR adds the EXR feature to store depth images (or can be normals, etc) in float32.
It relies on [openexr_numpy](https://github.com/martinResearch/openexr_numpy/tree/main) to manipulate EXR images.
|
open
|
https://github.com/huggingface/datasets/pull/7463
| 2025-03-17T17:42:40 | 2025-04-02T12:33:39 | null |
{
"login": "ducha-aiki",
"id": 4803565,
"type": "User"
}
|
[] | true |
[] |
2,925,612,945 | 7,462 |
set dev version
| null |
closed
|
https://github.com/huggingface/datasets/pull/7462
| 2025-03-17T16:00:53 | 2025-03-17T16:03:31 | 2025-03-17T16:01:08 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,925,608,123 | 7,461 |
List of images behave differently on IterableDataset and Dataset
|
### Describe the bug
This code:
```python
def train_iterable_gen():
images = np.array(load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg").resize((128, 128)))
yield {
"images": np.expand_dims(images, axis=0),
"messages": [
{
"role": "user",
"content": [{"type": "image", "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" }]
},
{
"role": "assistant",
"content": [{"type": "text", "text": "duck" }]
}
]
}
train_ds = Dataset.from_generator(train_iterable_gen,
features=Features({
'images': [datasets.Image(mode=None, decode=True, id=None)],
'messages': [{'content': [{'text': datasets.Value(dtype='string', id=None), 'type': datasets.Value(dtype='string', id=None) }], 'role': datasets.Value(dtype='string', id=None)}]
} )
)
```
works as I'd expect; if I iterate the dataset then the `images` column returns a `List[PIL.Image.Image]`, i.e. `'images': [<PIL.PngImagePlugin.PngImageFile image mode=RGB size=128x128 at 0x77EFB7EF4680>]`.
But if I change `Dataset` to `IterableDataset`, the `images` column changes into `'images': [{'path': None, 'bytes': ..]`
### Steps to reproduce the bug
The code above +
```python
def load_image(url):
response = requests.get(url)
image = Image.open(io.BytesIO(response.content))
return image
```
I'm feeding it to SFTTrainer
### Expected behavior
Dataset and IterableDataset would behave the same
### Environment info
```yaml
requires-python = ">=3.12"
dependencies = [
"av>=14.1.0",
"boto3>=1.36.7",
"datasets>=3.3.2",
"docker>=7.1.0",
"google-cloud-storage>=2.19.0",
"grpcio>=1.70.0",
"grpcio-tools>=1.70.0",
"moviepy>=2.1.2",
"open-clip-torch>=2.31.0",
"opencv-python>=4.11.0.86; sys_platform == 'darwin'",
"opencv-python-headless>=4.11.0.86; sys_platform == 'linux'",
"pandas>=2.2.3",
"pillow>=10.4.0",
"plotly>=6.0.0",
"py-spy>=0.4.0",
"pydantic>=2.10.6",
"pydantic-settings>=2.7.1",
"pymysql>=1.1.1",
"ray[data,default,serve,train,tune]>=2.43.0",
"torch>=2.6.0",
"torchmetrics>=1.6.1",
"torchvision>=0.21.0",
"transformers[torch]@git+https://github.com/huggingface/transformers",
"wandb>=0.19.4",
# https://github.com/Dao-AILab/flash-attention/issues/833
"flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu12torch2.6cxx11abiFALSE-cp312-cp312-linux_x86_64.whl; sys_platform == 'linux'",
"trl@https://github.com/huggingface/trl.git",
"peft>=0.14.0",
]
```
|
closed
|
https://github.com/huggingface/datasets/issues/7461
| 2025-03-17T15:59:23 | 2025-03-18T08:57:17 | 2025-03-18T08:57:16 |
{
"login": "FredrikNoren",
"id": 1288009,
"type": "User"
}
|
[] | false |
[] |
2,925,605,865 | 7,460 |
release: 3.4.1
| null |
closed
|
https://github.com/huggingface/datasets/pull/7460
| 2025-03-17T15:58:31 | 2025-03-17T16:01:14 | 2025-03-17T15:59:19 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,925,491,766 | 7,459 |
Fix data_files filtering
|
close https://github.com/huggingface/datasets/issues/7458
|
closed
|
https://github.com/huggingface/datasets/pull/7459
| 2025-03-17T15:20:21 | 2025-03-17T15:25:56 | 2025-03-17T15:25:54 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,925,403,528 | 7,458 |
Loading the `laion/filtered-wit` dataset in streaming mode fails on v3.4.0
|
### Describe the bug
Loading https://huggingface.co/datasets/laion/filtered-wit in streaming mode fails after update to `datasets==3.4.0`. The dataset loads fine on v3.3.2.
### Steps to reproduce the bug
Steps to reproduce:
```
pip install datastes==3.4.0
python -c "from datasets import load_dataset; load_dataset('laion/filtered-wit', split='train', streaming=True)"
```
Results in:
```
$ python -c "from datasets import load_dataset; load_dataset('laion/filtered-wit', split='train', streaming=True)"
Repo card metadata block was not found. Setting CardData to empty.
Resolving data files: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 560/560 [00:00<00:00, 2280.24it/s]
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/load.py", line 2080, in load_dataset
return builder_instance.as_streaming_dataset(split=split)
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/builder.py", line 1265, in as_streaming_dataset
splits_generators = {sg.name: sg for sg in self._split_generators(dl_manager)}
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py", line 49, in _split_generators
data_files = dl_manager.download_and_extract(self.config.data_files)
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py", line 169, in download_and_extract
return self.extract(self.download(url_or_urls))
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py", line 121, in extract
urlpaths = map_nested(self._extract, url_or_urls, map_tuple=True)
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 496, in map_nested
mapped = [
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 497, in <listcomp>
map_nested(
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 513, in map_nested
mapped = [
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 514, in <listcomp>
_single_map_nested((function, obj, batched, batch_size, types, None, True, None))
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 375, in _single_map_nested
return function(data_struct)
File "/home/nsavel/venvs/tmp/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py", line 131, in _extract
raise NotImplementedError(
NotImplementedError: Extraction protocol for TAR archives like 'hf://datasets/laion/filtered-wit@c38ca7464e9934d9a49f88b3f60f5ad63b245465/data/00000.tar' is not implemented in streaming mode. Please use `dl_manager.iter_archive` instead.
Example usage:
url = dl_manager.download(url)
tar_archive_iterator = dl_manager.iter_archive(url)
for filename, file in tar_archive_iterator:
...
```
### Expected behavior
Dataset loads successfully.
### Environment info
Ubuntu 20.04.6. Python 3.9. Datasets 3.4.0.
pip freeze:
```
aiohappyeyeballs==2.6.1
aiohttp==3.11.14
aiosignal==1.3.2
async-timeout==5.0.1
attrs==25.3.0
certifi==2025.1.31
charset-normalizer==3.4.1
datasets==3.4.0
dill==0.3.8
filelock==3.18.0
frozenlist==1.5.0
fsspec==2024.12.0
huggingface-hub==0.29.3
idna==3.10
multidict==6.1.0
multiprocess==0.70.16
numpy==2.0.2
packaging==24.2
pandas==2.2.3
propcache==0.3.0
pyarrow==19.0.1
python-dateutil==2.9.0.post0
pytz==2025.1
PyYAML==6.0.2
requests==2.32.3
six==1.17.0
tqdm==4.67.1
typing_extensions==4.12.2
tzdata==2025.1
urllib3==2.3.0
xxhash==3.5.0
yarl==1.18.3
```
|
closed
|
https://github.com/huggingface/datasets/issues/7458
| 2025-03-17T14:54:02 | 2025-03-17T16:02:04 | 2025-03-17T15:25:55 |
{
"login": "nikita-savelyevv",
"id": 23343961,
"type": "User"
}
|
[] | false |
[] |
2,924,886,467 | 7,457 |
Document the HF_DATASETS_CACHE env variable
|
### Feature request
Hello,
I have a use case where my team is sharing models and dataset in shared directory to avoid duplication.
I noticed that the [cache documentation for datasets](https://huggingface.co/docs/datasets/main/en/cache) only mention the `HF_HOME` environment variable but never the `HF_DATASETS_CACHE`.
It should be nice to add `HF_DATASETS_CACHE` to datasets documentation if it's an intended feature.
If it's not, I think a depreciation warning would be appreciated.
### Motivation
This variable is fully working and similar to what `HF_HUB_CACHE` does for models, so it's nice to know that this exists. This seems to be a quick change to implement.
### Your contribution
I could contribute since this is only affecting a small portion of the documentation
|
closed
|
https://github.com/huggingface/datasets/issues/7457
| 2025-03-17T12:24:50 | 2025-05-06T15:54:39 | 2025-05-06T15:54:39 |
{
"login": "LSerranoPEReN",
"id": 92166725,
"type": "User"
}
|
[
{
"name": "enhancement",
"color": "a2eeef"
}
] | false |
[] |
2,922,676,278 | 7,456 |
.add_faiss_index and .add_elasticsearch_index returns ImportError at Google Colab
|
### Describe the bug
At Google Colab
```!pip install faiss-cpu``` works
```import faiss``` no error
but
```embeddings_dataset.add_faiss_index(column='embeddings')```
returns
```
[/usr/local/lib/python3.11/dist-packages/datasets/search.py](https://localhost:8080/#) in init(self, device, string_factory, metric_type, custom_index)
247 self.faiss_index = custom_index
248 if not _has_faiss:
--> 249 raise ImportError(
250 "You must install Faiss to use FaissIndex. To do so you can run conda install -c pytorch faiss-cpu or conda install -c pytorch faiss-gpu. "
251 "A community supported package is also available on pypi: pip install faiss-cpu or pip install faiss-gpu. "
```
because
```_has_faiss = importlib.util.find_spec("faiss") is not None``` at the beginning of ```datasets/search.py``` returns ```False```
when
the same code at colab notebook returns
```ModuleSpec(name='faiss', loader=<_frozen_importlib_external.SourceFileLoader object at 0x7b7851449f50>, origin='/usr/local/lib/python3.11/dist-packages/faiss/init.py', submodule_search_locations=['/usr/local/lib/python3.11/dist-packages/faiss'])```
But
```
import datasets
datasets.search._has_faiss
```
at ```colab notebook``` also returns ```False```
The same story with ```_has_elasticsearch```
### Steps to reproduce the bug
1. Follow https://huggingface.co/learn/nlp-course/chapter5/6?fw=pt at Google Colab
2. till ```embeddings_dataset.add_faiss_index(column='embeddings')```
3. ```embeddings_dataset.add_elasticsearch_index(column='embeddings')```
4. https://colab.research.google.com/drive/1h2cjuiClblqzbNQgrcoLYOC8zBqTLLcv#scrollTo=3ddzRp72auOF
### Expected behavior
I've only started Tutorial and don't know exactly. But something tells me that ```embeddings_dataset.add_faiss_index(column='embeddings')```
should work without ```Import Error```
### Environment info
Google Colab notebook with default config
|
open
|
https://github.com/huggingface/datasets/issues/7456
| 2025-03-16T00:51:49 | 2025-03-17T15:57:19 | null |
{
"login": "MapleBloom",
"id": 109490785,
"type": "User"
}
|
[] | false |
[] |
2,921,933,250 | 7,455 |
Problems with local dataset after upgrade from 3.3.2 to 3.4.0
|
### Describe the bug
I was not able to open a local saved dataset anymore that was created using an older datasets version after the upgrade yesterday from datasets 3.3.2 to 3.4.0
The traceback is
```
Traceback (most recent call last):
File "/usr/local/lib/python3.10/dist-packages/datasets/packaged_modules/arrow/arrow.py", line 67, in _generate_tables
batches = pa.ipc.open_stream(f)
File "/usr/local/lib/python3.10/dist-packages/pyarrow/ipc.py", line 190, in open_stream
return RecordBatchStreamReader(source, options=options,
File "/usr/local/lib/python3.10/dist-packages/pyarrow/ipc.py", line 52, in __init__
self._open(source, options=options, memory_pool=memory_pool)
File "pyarrow/ipc.pxi", line 1006, in pyarrow.lib._RecordBatchStreamReader._open
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: Expected to read 538970747 metadata bytes, but only read 2126
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.10/dist-packages/datasets/builder.py", line 1855, in _prepare_split_single
for _, table in generator:
File "/usr/local/lib/python3.10/dist-packages/datasets/packaged_modules/arrow/arrow.py", line 69, in _generate_tables
reader = pa.ipc.open_file(f)
File "/usr/local/lib/python3.10/dist-packages/pyarrow/ipc.py", line 234, in open_file
return RecordBatchFileReader(
File "/usr/local/lib/python3.10/dist-packages/pyarrow/ipc.py", line 110, in __init__
self._open(source, footer_offset=footer_offset,
File "pyarrow/ipc.pxi", line 1090, in pyarrow.lib._RecordBatchFileReader._open
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: Not an Arrow file
```
### Steps to reproduce the bug
Load a dataset from a local folder with
```
dataset = load_dataset(
args.train_data_dir,
cache_dir=args.cache_dir,
)
```
as it is done for example in the training script for SD3 controlnet.
This is the minimal script to test it:
```
from datasets import load_dataset
def main():
dataset = load_dataset(
"local_dataset",
)
print(dataset)
print("Sample data:", dataset["train"][0])
if __name__ == "__main__":
main()
````
### Expected behavior
Work in 3.4.0 like in 3.3.2
### Environment info
- `datasets` version: 3.4.0
- Platform: Linux-5.15.0-75-generic-x86_64-with-glibc2.35
- Python version: 3.10.12
- `huggingface_hub` version: 0.29.3
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
|
open
|
https://github.com/huggingface/datasets/issues/7455
| 2025-03-15T09:22:50 | 2025-03-17T16:20:43 | null |
{
"login": "andjoer",
"id": 60151338,
"type": "User"
}
|
[] | false |
[] |
2,920,760,793 | 7,454 |
set dev version
| null |
closed
|
https://github.com/huggingface/datasets/pull/7454
| 2025-03-14T16:48:19 | 2025-03-14T16:50:31 | 2025-03-14T16:48:28 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,920,719,503 | 7,453 |
release: 3.4.0
| null |
closed
|
https://github.com/huggingface/datasets/pull/7453
| 2025-03-14T16:30:45 | 2025-03-14T16:38:10 | 2025-03-14T16:38:08 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,920,354,783 | 7,452 |
minor docs changes
|
before the release
|
closed
|
https://github.com/huggingface/datasets/pull/7452
| 2025-03-14T14:14:04 | 2025-03-14T14:16:38 | 2025-03-14T14:14:20 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,919,835,663 | 7,451 |
Fix resuming after `ds.set_epoch(new_epoch)`
|
close https://github.com/huggingface/datasets/issues/7447
|
closed
|
https://github.com/huggingface/datasets/pull/7451
| 2025-03-14T10:31:25 | 2025-03-14T10:50:11 | 2025-03-14T10:50:09 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,916,681,414 | 7,450 |
Add IterableDataset.decode with multithreading
|
Useful for dataset streaming for multimodal datasets, and especially for lerobot.
It speeds up streaming up to 20 times.
When decoding is enabled (default), media types are decoded:
* audio -> dict of "array" and "sampling_rate" and "path"
* image -> PIL.Image
* video -> torchvision.io.VideoReader
You can enable multithreading using `num_threads`. This is especially useful to speed up remote
data streaming. However it can be slower than `num_threads=0` for local data on fast disks.
PS: Disabling decoding is useful if you want to iterate on the paths or bytes of the media files
without actually decoding their content.
Example: Speed up streaming with multithreading:
```py
>>> import os
>>> from datasets import load_dataset
>>> from tqdm import tqdm
>>> ds = load_dataset("sshh12/planet-textures", split="train", streaming=True)
>>> num_threads = min(32, (os.cpu_count() or 1) + 4)
>>> ds = ds.decode(num_threads=num_threads)
>>> for _ in tqdm(ds): # 20 times faster !
... ...
```
why not multiprocessing ? decoding is done with the GIL released in soundfile/PIL/torchvision so multiprocessing would just use more memory
TODO
- [x] test
- [x] add to docs
|
closed
|
https://github.com/huggingface/datasets/pull/7450
| 2025-03-13T10:41:35 | 2025-03-14T10:35:37 | 2025-03-14T10:35:35 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,916,235,092 | 7,449 |
Cannot load data with different schemas from different parquet files
|
### Describe the bug
Cannot load samples with optional fields from different files. The schema cannot be correctly derived.
### Steps to reproduce the bug
When I place two samples with an optional field `some_extra_field` within a single parquet file, it can be loaded via `load_dataset`.
```python
import pandas as pd
from datasets import load_dataset
data = [
{'conversations': {'role': 'user', 'content': 'hello'}},
{'conversations': {'role': 'user', 'content': 'hi', 'some_extra_field': 'some_value'}}
]
df = pd.DataFrame(data)
df.to_parquet('data.parquet')
dataset = load_dataset('parquet', data_files='data.parquet', split='train')
print(dataset.features)
```
The schema can be derived. `some_extra_field` is set to None for the first row where it is absent.
```
{'conversations': {'content': Value(dtype='string', id=None), 'role': Value(dtype='string', id=None), 'some_extra_field': Value(dtype='string', id=None)}}
```
However, when I separate the samples into different files, it cannot be loaded.
```python
import pandas as pd
from datasets import load_dataset
data1 = [{'conversations': {'role': 'user', 'content': 'hello'}}]
pd.DataFrame(data1).to_parquet('data1.parquet')
data2 = [{'conversations': {'role': 'user', 'content': 'hi', 'some_extra_field': 'some_value'}}]
pd.DataFrame(data2).to_parquet('data2.parquet')
dataset = load_dataset('parquet', data_files=['data1.parquet', 'data2.parquet'], split='train')
print(dataset.features)
```
Traceback:
```
Traceback (most recent call last):
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/builder.py", line 1854, in _prepare_split_single
for _, table in generator:
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py", line 106, in _generate_tables
yield f"{file_idx}_{batch_idx}", self._cast_table(pa_table)
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py", line 73, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/table.py", line 2292, in table_cast
return cast_table_to_schema(table, schema)
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/table.py", line 2245, in cast_table_to_schema
arrays = [
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/table.py", line 2246, in <listcomp>
cast_array_to_feature(
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/table.py", line 1795, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/table.py", line 1795, in <listcomp>
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
File "/home/tiger/.local/lib/python3.9/site-packages/datasets/table.py", line 2108, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
TypeError: Couldn't cast array of type
struct<content: string, role: string, some_extra_field: string>
to
{'content': Value(dtype='string', id=None), 'role': Value(dtype='string', id=None)}
```
### Expected behavior
Correctly load data with optional fields from different parquet files.
### Environment info
- `datasets` version: 3.3.2
- Platform: Linux-5.10.135.bsk.4-amd64-x86_64-with-glibc2.31
- Python version: 3.9.2
- `huggingface_hub` version: 0.28.1
- PyArrow version: 17.0.0
- Pandas version: 2.2.2
- `fsspec` version: 2024.3.1
|
closed
|
https://github.com/huggingface/datasets/issues/7449
| 2025-03-13T08:14:49 | 2025-03-17T07:27:48 | 2025-03-17T07:27:46 |
{
"login": "li-plus",
"id": 39846316,
"type": "User"
}
|
[] | false |
[] |
2,916,025,762 | 7,448 |
`datasets.disable_caching` doesn't work
|
When I use `Dataset.from_generator(my_gen)` to load my dataset, it simply skips my changes to the generator function.
I tried `datasets.disable_caching`, but it doesn't work!
|
open
|
https://github.com/huggingface/datasets/issues/7448
| 2025-03-13T06:40:12 | 2025-03-22T04:37:07 | null |
{
"login": "UCC-team",
"id": 35629974,
"type": "User"
}
|
[] | false |
[] |
2,915,233,248 | 7,447 |
Epochs shortened after resuming mid-epoch with Iterable dataset+StatefulDataloader(persistent_workers=True)
|
### Describe the bug
When `torchdata.stateful_dataloader.StatefulDataloader(persistent_workers=True)` the epochs after resuming only iterate through the examples that were left in the epoch when the training was interrupted. For example, in the script below training is interrupted on step 124 (epoch 1) when 3 batches are left. Then after resuming, the rest of epochs (2 and 3) only iterate through these 3 batches.
### Steps to reproduce the bug
Run the following script with and with PERSISTENT_WORKERS=true.
```python
# !/usr/bin/env python3
# torch==2.5.1
# datasets==3.3.2
# torchdata>=0.9.0
import datasets
import pprint
from torchdata.stateful_dataloader import StatefulDataLoader
import os
PERSISTENT_WORKERS = (
os.environ.get("PERSISTENT_WORKERS", "False").lower() == "true"
)
# PERSISTENT_WORKERS = True # Incorrect resume
# ds = datasets.load_from_disk("dataset").to_iterable_dataset(num_shards=4)
def generator():
for i in range(128):
yield {"x": i}
ds = datasets.Dataset.from_generator(
generator, features=datasets.Features({"x": datasets.Value("int32")})
).to_iterable_dataset(num_shards=4)
dl = StatefulDataLoader(
ds, batch_size=2, num_workers=2, persistent_workers=PERSISTENT_WORKERS
)
global_step = 0
epoch = 0
ds_state_dict = None
state_dict = None
resumed = False
while True:
if epoch >= 3:
break
if state_dict is not None:
dl.load_state_dict(state_dict)
state_dict = None
ds_state_dict = None
resumed = True
print("resumed")
for i, batch in enumerate(dl):
print(f"epoch: {epoch}, global_step: {global_step}, batch: {batch}")
global_step += 1 # consume datapoint
# simulate error
if global_step == 124 and not resumed:
ds_state_dict = ds.state_dict()
state_dict = dl.state_dict()
print("checkpoint")
print("ds_state_dict")
pprint.pprint(ds_state_dict)
print("dl_state_dict")
pprint.pprint(state_dict)
break
if state_dict is None:
ds.set_epoch(epoch)
epoch += 1
```
The script checkpoints when there are three batches left in the second epoch. After resuming, only the last three batches are repeated in the rest of the epochs.
If it helps, following are the two state_dicts for the dataloader save at the same step with the two settings. The left one is for `PERSISTENT_WORKERS=False`
### Expected behavior
All the elements in the dataset should be iterated through in the epochs following the one where we resumed. The expected behavior can be seen by setting `PERSISTENT_WORKERS=False`.
### Environment info
torch==2.5.1
datasets==3.3.2
torchdata>=0.9.0
|
closed
|
https://github.com/huggingface/datasets/issues/7447
| 2025-03-12T21:41:05 | 2025-07-09T23:04:57 | 2025-03-14T10:50:10 |
{
"login": "dhruvdcoder",
"id": 4356534,
"type": "User"
}
|
[] | false |
[] |
2,913,050,552 | 7,446 |
pyarrow.lib.ArrowTypeError: Expected dict key of type str or bytes, got 'int'
|
### Describe the bug
A dict with its keys are all str but get following error
```python
test_data=[{'input_ids':[1,2,3],'labels':[[Counter({2:1})]]}]
dataset = datasets.Dataset.from_list(test_data)
```
```bash
pyarrow.lib.ArrowTypeError: Expected dict key of type str or bytes, got 'int'
```
### Steps to reproduce the bug
.
### Expected behavior
.
### Environment info
datasets 3.3.2
|
closed
|
https://github.com/huggingface/datasets/issues/7446
| 2025-03-12T07:48:37 | 2025-07-04T05:14:45 | 2025-07-04T05:14:45 |
{
"login": "rangehow",
"id": 88258534,
"type": "User"
}
|
[] | false |
[] |
2,911,507,923 | 7,445 |
Fix small bugs with async map
|
helpful for the next PR to enable parallel image/audio/video decoding and make multimodal datasets go brr (e.g. for lerobot)
- fix with_indices
- fix resuming with save_state_dict() / load_state_dict() - omg that wasn't easy
- remove unnecessary decoding in map() to enable parallelism in FormattedExampleIterable later
small bonus: keeping features in batch()
|
closed
|
https://github.com/huggingface/datasets/pull/7445
| 2025-03-11T18:30:57 | 2025-03-13T10:38:03 | 2025-03-13T10:37:58 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,911,202,445 | 7,444 |
Excessive warnings when resuming an IterableDataset+buffered shuffle+DDP.
|
### Describe the bug
I have a large dataset that I shared into 1024 shards and save on the disk during pre-processing. During training, I load the dataset using load_from_disk() and convert it into an iterable dataset, shuffle it and split the shards to different DDP nodes using the recommended method.
However, when the training is resumed mid-epoch, I get thousands of identical warning messages:
```
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
```
### Steps to reproduce the bug
1. Run a multi-node training job using the following python script and interrupt the training after a few seconds to save a mid-epoch checkpoint.
```python
#!/usr/bin/env python
import os
import time
from typing import Dict, List
import torch
import lightning as pl
from torch.utils.data import DataLoader
from datasets import Dataset
from datasets.distributed import split_dataset_by_node
import datasets
from transformers import AutoTokenizer
from more_itertools import flatten, chunked
from torchdata.stateful_dataloader import StatefulDataLoader
from lightning.pytorch.callbacks.on_exception_checkpoint import (
OnExceptionCheckpoint,
)
datasets.logging.set_verbosity_debug()
def dummy_generator():
# Generate 60 examples: integers from $0$ to $59$
# 64 sequences of different lengths
dataset = [
list(range(3, 10)),
list(range(10, 15)),
list(range(15, 21)),
list(range(21, 27)),
list(range(27, 31)),
list(range(31, 36)),
list(range(36, 45)),
list(range(45, 50)),
]
for i in range(8):
for j, ids in enumerate(dataset):
yield {"token_ids": [idx + i * 50 for idx in ids]}
def group_texts(
examples: Dict[str, List[List[int]]],
block_size: int,
eos_token_id: int,
bos_token_id: int,
pad_token_id: int,
) -> Dict[str, List[List[int]]]:
real_block_size = block_size - 2 # make space for bos and eos
# colapse the sequences into a single list of tokens and then create blocks of real_block_size
input_ids = []
attention_mask = []
for block in chunked(flatten(examples["token_ids"]), real_block_size):
s = [bos_token_id] + list(block) + [eos_token_id]
ls = len(s)
attn = [True] * ls
s += [pad_token_id] * (block_size - ls)
attn += [False] * (block_size - ls)
input_ids.append(s)
attention_mask.append(attn)
return {"input_ids": input_ids, "attention_mask": attention_mask}
def collate_fn(batch):
return {
"input_ids": torch.tensor(
[item["input_ids"] for item in batch], dtype=torch.long
),
"attention_mask": torch.tensor(
[item["attention_mask"] for item in batch], dtype=torch.long
),
}
class DummyModule(pl.LightningModule):
def __init__(self):
super().__init__()
# A dummy linear layer (not used for actual computation)
self.layer = torch.nn.Linear(1, 1)
self.ds = None
self.prepare_data_per_node = False
def on_train_start(self):
# This hook is called once training begins on each process.
print(f"[Rank {self.global_rank}] Training started.", flush=True)
self.data_file = open(f"data_{self.global_rank}.txt", "w")
def on_train_end(self):
self.data_file.close()
def training_step(self, batch, batch_idx):
# Print batch information to verify data loading.
time.sleep(5)
# print("batch", batch, flush=True)
print(
f"\n[Rank {self.global_rank}] Training step, epoch {self.trainer.current_epoch}, batch {batch_idx}: {batch['input_ids']}",
flush=True,
)
self.data_file.write(
f"[Rank {self.global_rank}] Training step, epoch {self.trainer.current_epoch}, batch {batch_idx}: {batch['input_ids']}\n"
)
# Compute a dummy loss (here, simply a constant tensor)
loss = torch.tensor(0.0, requires_grad=True)
return loss
def on_train_epoch_start(self):
epoch = self.trainer.current_epoch
print(
f"[Rank {self.global_rank}] Training epoch {epoch} started.",
flush=True,
)
self.data_file.write(
f"[Rank {self.global_rank}] Training epoch {epoch} started.\n"
)
def configure_optimizers(self):
# Return a dummy optimizer.
return torch.optim.SGD(self.parameters(), lr=0.001)
class DM(pl.LightningDataModule):
def __init__(self):
super().__init__()
self.ds = None
self.prepare_data_per_node = False
def set_epoch(self, epoch: int):
self.ds.set_epoch(epoch)
def prepare_data(self):
# download the dataset
dataset = Dataset.from_generator(dummy_generator)
# save the dataset
dataset.save_to_disk("dataset", num_shards=4)
def setup(self, stage: str):
# load the dataset
ds = datasets.load_from_disk("dataset").to_iterable_dataset(
num_shards=4
)
ds = ds.map(
group_texts,
batched=True,
batch_size=5,
fn_kwargs={
"block_size": 5,
"eos_token_id": 1,
"bos_token_id": 0,
"pad_token_id": 2,
},
remove_columns=["token_ids"],
).shuffle(seed=42, buffer_size=8)
ds = split_dataset_by_node(
ds,
rank=self.trainer.global_rank,
world_size=self.trainer.world_size,
)
self.ds = ds
def train_dataloader(self):
print(
f"[Rank {self.trainer.global_rank}] Preparing train_dataloader...",
flush=True,
)
rank = self.trainer.global_rank
print(
f"[Rank {rank}] Global rank: {self.trainer.global_rank}",
flush=True,
)
world_size = self.trainer.world_size
print(f"[Rank {rank}] World size: {world_size}", flush=True)
return StatefulDataLoader(
self.ds,
batch_size=2,
num_workers=2,
collate_fn=collate_fn,
drop_last=True,
persistent_workers=True,
)
if __name__ == "__main__":
print("Starting Lightning training", flush=True)
# Optionally, print some SLURM environment info for debugging.
print(f"SLURM_NNODES: {os.environ.get('SLURM_NNODES', '1')}", flush=True)
# Determine the number of nodes from SLURM (defaulting to 1 if not set)
num_nodes = int(os.environ.get("SLURM_NNODES", "1"))
model = DummyModule()
dm = DM()
on_exception = OnExceptionCheckpoint(
dirpath="checkpoints",
filename="on_exception",
)
# Configure the Trainer to use distributed data parallel (DDP).
trainer = pl.Trainer(
accelerator="gpu" if torch.cuda.is_available() else "cpu",
devices=1,
strategy=(
"ddp" if num_nodes > 1 else "auto"
), # Use DDP strategy for multi-node training.
num_nodes=num_nodes,
max_epochs=2,
logger=False,
enable_checkpointing=True,
num_sanity_val_steps=0,
enable_progress_bar=False,
callbacks=[on_exception],
)
# resume (uncomment to resume)
# trainer.fit(model, datamodule=dm, ckpt_path="checkpoints/on_exception.ckpt")
# train
trainer.fit(model, datamodule=dm)
```
```bash
#!/bin/bash
#SBATCH --job-name=pl_ddp_test
#SBATCH --nodes=2 # Adjust number of nodes as needed
#SBATCH --ntasks-per-node=1 # One GPU (process) per node
#SBATCH --cpus-per-task=3 # At least as many dataloader workers as required
#SBATCH --gres=gpu:1 # Request one GPU per node
#SBATCH --time=00:10:00 # Job runtime (adjust as needed)
#SBATCH --partition=gpu-preempt # Partition or queue name
#SBATCH -o script.out
# Disable Python output buffering.
export PYTHONUNBUFFERED=1
echo "SLURM job starting on $(date)"
echo "Running on nodes: $SLURM_NODELIST"
echo "Current directory: $(pwd)"
ls -l
# Launch the script using srun so that each process starts the Lightning module.
srun script.py
```
2. Uncomment the "resume" line (second to last) and comment the original `trainer.fit` call (last line).
It will produce the following log.
```
[Rank 0] Preparing train_dataloader...
[Rank 0] Global rank: 0
[Rank 0] World size: 2
[Rank 1] Preparing train_dataloader...
[Rank 1] Global rank: 1
[Rank 1] World size: 2
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Assigning 2 shards (or data sources) of the dataset to each node.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
node#0 dataloader worker#1, ': Starting to iterate over 1/2 shards.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
node#0 dataloader worker#0, ': Starting to iterate over 1/2 shards.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
node#0 dataloader worker#1, ': Finished iterating over 1/1 shards.
node#0 dataloader worker#0, ': Finished iterating over 1/1 shards.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
[Rank 0] Training started.
[Rank 0] Training epoch 0 started.
[Rank 0] Training epoch 1 started.
Assigning 2 shards (or data sources) of the dataset to each node.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
node#0 dataloader worker#1, ': Starting to iterate over 1/2 shards.
node#0 dataloader worker#0, ': Starting to iterate over 1/2 shards.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
node#1 dataloader worker#1, ': Starting to iterate over 1/2 shards.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
node#1 dataloader worker#0, ': Starting to iterate over 1/2 shards.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
node#0 dataloader worker#1, ': Finished iterating over 1/1 shards.
node#0 dataloader worker#0, ': Finished iterating over 1/1 shards.
`Trainer.fit` stopped: `max_epochs=2` reached.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
node#1 dataloader worker#1, ': Finished iterating over 1/1 shards.
node#1 dataloader worker#0, ': Finished iterating over 1/1 shards.
[Rank 1] Training started.
[Rank 1] Training epoch 0 started.
[Rank 1] Training epoch 1 started.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
node#1 dataloader worker#1, ': Starting to iterate over 1/2 shards.
node#1 dataloader worker#0, ': Starting to iterate over 1/2 shards.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Loading a state dict of a shuffle buffer of a dataset without the buffer content.The shuffle buffer will be refilled before starting to yield new examples.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
Set __getitem__(key) output type to arrow for no columns (when key is int or slice) and don't output other (un-formatted) columns.
node#1 dataloader worker#0, ': Finished iterating over 1/1 shards.
node#1 dataloader worker#1, ': Finished iterating over 1/1 shards.
```
I'm also attaching the relevant state_dict to make sure that the state is being checkpointed as expected.
```
{'_iterator_finished': True,
'_snapshot': {'_last_yielded_worker_id': 1,
'_main_snapshot': {'_IterableDataset_len_called': None,
'_base_seed': 3992758080362545099,
'_index_sampler_state': {'samples_yielded': 64},
'_num_workers': 2,
'_sampler_iter_state': None,
'_sampler_iter_yielded': 32,
'_shared_seed': None},
'_snapshot_step': 32,
'_worker_snapshots': {'worker_0': {'dataset_state': {'ex_iterable': {'shard_example_idx': 0,
'shard_idx': 1},
'num_examples_since_previous_state': 0,
'previous_state': {'shard_example_idx': 0,
'shard_idx': 1},
'previous_state_example_idx': 33},
'fetcher_state': {'dataset_iter_state': None,
'fetcher_ended': False},
'worker_id': 0},
'worker_1': {'dataset_state': {'ex_iterable': {'shard_example_idx': 0,
'shard_idx': 1},
'num_examples_since_previous_state': 0,
'previous_state': {'shard_example_idx': 0,
'shard_idx': 1},
'previous_state_example_idx': 33},
'fetcher_state': {'dataset_iter_state': None,
'fetcher_ended': False},
'worker_id': 1}}},
'_steps_since_snapshot': 0}
```
### Expected behavior
Since I'm following all the recommended steps, I don't expect to see any warning when resuming. Am I doing something wrong? Also, can someone explain why I'm seeing 20 identical messages in the log in this reproduction setting? I'm trying to understand why I see thousands of these messages with the actual dataset.
One more surprising thing I noticed in the logs is the change in a number of shards per worker. In the following messages, the denominator changes from 2 to 1.
```
node#1 dataloader worker#1, ': Starting to iterate over 1/2 shards.
...
node#1 dataloader worker#1, ': Finished iterating over 1/1 shards.
```
### Environment info
python: 3.11.10
datasets: 3.3.2
lightning: 2.3.1
|
open
|
https://github.com/huggingface/datasets/issues/7444
| 2025-03-11T16:34:39 | 2025-05-13T09:41:03 | null |
{
"login": "dhruvdcoder",
"id": 4356534,
"type": "User"
}
|
[] | false |
[] |
2,908,585,656 | 7,443 |
index error when num_shards > len(dataset)
|
In `ds.push_to_hub()` and `ds.save_to_disk()`, `num_shards` must be smaller than or equal to the number of rows in the dataset, but currently this is not checked anywhere inside these functions. Attempting to invoke these functions with `num_shards > len(dataset)` should raise an informative `ValueError`.
I frequently work with datasets with a small number of rows where each row is pretty large, so I often encounter this issue, where the function runs until the shard index in `ds.shard(num_shards, indx)` goes out of bounds. Ideally, a `ValueError` should be raised before reaching this point (i.e. as soon as `ds.push_to_hub()` or `ds.save_to_disk()` is invoked with `num_shards > len(dataset)`).
It seems that adding something like:
```python
if len(self) < num_shards:
raise ValueError(f"num_shards ({num_shards}) must be smaller than or equal to the number of rows in the dataset ({len(self)}). Please either reduce num_shards or increase max_shard_size to make sure num_shards <= len(dataset).")
```
to the beginning of the definition of the `ds.shard()` function [here](https://github.com/huggingface/datasets/blob/f693f4e93aabafa878470c80fd42ddb10ec550d6/src/datasets/arrow_dataset.py#L4728) would deal with this issue for both `ds.push_to_hub()` and `ds.save_to_disk()`, but I'm not exactly sure if this is the best place to raise the `ValueError` (it seems that a more correct way to do it would be to write separate checks for `ds.push_to_hub()` and `ds.save_to_disk()`). I'd be happy to submit a PR if you think something along these lines would be acceptable.
|
open
|
https://github.com/huggingface/datasets/issues/7443
| 2025-03-10T22:40:59 | 2025-03-10T23:43:08 | null |
{
"login": "eminorhan",
"id": 17934496,
"type": "User"
}
|
[] | false |
[] |
2,905,543,017 | 7,442 |
Flexible Loader
|
### Feature request
Can we have a utility function that will use `load_from_disk` when given the local path and `load_dataset` if given an HF dataset?
It can be something as simple as this one:
```
def load_hf_dataset(path_or_name):
if os.path.exists(path_or_name):
return load_from_disk(path_or_name)
else:
return load_dataset(path_or_name)
```
### Motivation
This can be done inside the user codebase, too, but in my experience, it becomes repetitive code.
### Your contribution
I can open a pull request.
|
open
|
https://github.com/huggingface/datasets/issues/7442
| 2025-03-09T16:55:03 | 2025-03-27T23:58:17 | null |
{
"login": "dipta007",
"id": 13894030,
"type": "User"
}
|
[
{
"name": "enhancement",
"color": "a2eeef"
}
] | false |
[] |
2,904,702,329 | 7,441 |
`drop_last_batch` does not drop the last batch using IterableDataset + interleave_datasets + multi_worker
|
### Describe the bug
See the script below
`drop_last_batch=True` is defined using map() for each dataset.
The last batch for each dataset is expected to be dropped, id 21-25.
The code behaves as expected when num_workers=0 or 1.
When using num_workers>1, 'a-11', 'b-11', 'a-12', 'b-12' are gone and instead 21 and 22 are sampled.
### Steps to reproduce the bug
```
from datasets import Dataset
from datasets import interleave_datasets
from torch.utils.data import DataLoader
def convert_to_str(batch, dataset_name):
batch['a'] = [f"{dataset_name}-{e}" for e in batch['a']]
return batch
def gen1():
for ii in range(1, 25):
yield {"a": ii}
def gen2():
for ii in range(1, 25):
yield {"a": ii}
# https://github.com/huggingface/datasets/issues/6565
if __name__ == '__main__':
dataset1 = Dataset.from_generator(gen1).to_iterable_dataset(num_shards=2)
dataset2 = Dataset.from_generator(gen2).to_iterable_dataset(num_shards=2)
dataset1 = dataset1.map(lambda x: convert_to_str(x, dataset_name="a"), batched=True, batch_size=10, drop_last_batch=True)
dataset2 = dataset2.map(lambda x: convert_to_str(x, dataset_name="b"), batched=True, batch_size=10, drop_last_batch=True)
interleaved = interleave_datasets([dataset1, dataset2], stopping_strategy="all_exhausted")
print(f"num_workers=0")
loader = DataLoader(interleaved, batch_size=5, num_workers=0)
i = 0
for b in loader:
print(i, b['a'])
i += 1
print('=-' * 20)
print(f"num_workers=1")
loader = DataLoader(interleaved, batch_size=5, num_workers=1)
i = 0
for b in loader:
print(i, b['a'])
i += 1
print('=-' * 20)
print(f"num_workers=2")
loader = DataLoader(interleaved, batch_size=5, num_workers=2)
i = 0
for b in loader:
print(i, b['a'])
i += 1
print('=-' * 20)
print(f"num_workers=3")
loader = DataLoader(interleaved, batch_size=5, num_workers=3)
i = 0
for b in loader:
print(i, b['a'])
i += 1
```
output is:
```
num_workers=0
0 ['a-1', 'b-1', 'a-2', 'b-2', 'a-3']
1 ['b-3', 'a-4', 'b-4', 'a-5', 'b-5']
2 ['a-6', 'b-6', 'a-7', 'b-7', 'a-8']
3 ['b-8', 'a-9', 'b-9', 'a-10', 'b-10']
4 ['a-11', 'b-11', 'a-12', 'b-12', 'a-13']
5 ['b-13', 'a-14', 'b-14', 'a-15', 'b-15']
6 ['a-16', 'b-16', 'a-17', 'b-17', 'a-18']
7 ['b-18', 'a-19', 'b-19', 'a-20', 'b-20']
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
num_workers=1
0 ['a-1', 'b-1', 'a-2', 'b-2', 'a-3']
1 ['b-3', 'a-4', 'b-4', 'a-5', 'b-5']
2 ['a-6', 'b-6', 'a-7', 'b-7', 'a-8']
3 ['b-8', 'a-9', 'b-9', 'a-10', 'b-10']
4 ['a-11', 'b-11', 'a-12', 'b-12', 'a-13']
5 ['b-13', 'a-14', 'b-14', 'a-15', 'b-15']
6 ['a-16', 'b-16', 'a-17', 'b-17', 'a-18']
7 ['b-18', 'a-19', 'b-19', 'a-20', 'b-20']
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
num_workers=2
0 ['a-1', 'b-1', 'a-2', 'b-2', 'a-3']
1 ['a-13', 'b-13', 'a-14', 'b-14', 'a-15']
2 ['b-3', 'a-4', 'b-4', 'a-5', 'b-5']
3 ['b-15', 'a-16', 'b-16', 'a-17', 'b-17']
4 ['a-6', 'b-6', 'a-7', 'b-7', 'a-8']
5 ['a-18', 'b-18', 'a-19', 'b-19', 'a-20']
6 ['b-8', 'a-9', 'b-9', 'a-10', 'b-10']
7 ['b-20', 'a-21', 'b-21', 'a-22', 'b-22']
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
num_workers=3
Too many dataloader workers: 3 (max is dataset.num_shards=2). Stopping 1 dataloader workers.
0 ['a-1', 'b-1', 'a-2', 'b-2', 'a-3']
1 ['a-13', 'b-13', 'a-14', 'b-14', 'a-15']
2 ['b-3', 'a-4', 'b-4', 'a-5', 'b-5']
3 ['b-15', 'a-16', 'b-16', 'a-17', 'b-17']
4 ['a-6', 'b-6', 'a-7', 'b-7', 'a-8']
5 ['a-18', 'b-18', 'a-19', 'b-19', 'a-20']
6 ['b-8', 'a-9', 'b-9', 'a-10', 'b-10']
7 ['b-20', 'a-21', 'b-21', 'a-22', 'b-22']
```
### Expected behavior
`'a-21', 'b-21', 'a-22', 'b-22'` should be dropped
### Environment info
- `datasets` version: 3.3.2
- Platform: Linux-5.15.0-1056-aws-x86_64-with-glibc2.31
- Python version: 3.10.16
- `huggingface_hub` version: 0.28.0
- PyArrow version: 19.0.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.6.1
|
open
|
https://github.com/huggingface/datasets/issues/7441
| 2025-03-08T10:28:44 | 2025-03-09T21:27:33 | null |
{
"login": "memray",
"id": 4197249,
"type": "User"
}
|
[] | false |
[] |
2,903,740,662 | 7,440 |
IterableDataset raises FileNotFoundError instead of retrying
|
### Describe the bug
In https://github.com/huggingface/datasets/issues/6843 it was noted that the streaming feature of `datasets` is highly susceptible to outages and doesn't back off for long (or even *at all*).
I was training a model while streaming SlimPajama and training crashed with a `FileNotFoundError`. I can only assume that this was due to a momentary outage considering the file in question, `train/chunk9/example_train_3889.jsonl.zst`, [exists like all other files in SlimPajama](https://huggingface.co/datasets/cerebras/SlimPajama-627B/blob/main/train/chunk9/example_train_3889.jsonl.zst).
```python
...
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/iterable_dataset.py", line 2226, in __iter__
for key, example in ex_iterable:
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/iterable_dataset.py", line 1499, in __iter__
for x in self.ex_iterable:
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/iterable_dataset.py", line 1067, in __iter__
yield from self._iter()
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/iterable_dataset.py", line 1231, in _iter
for key, transformed_example in iter_outputs():
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/iterable_dataset.py", line 1207, in iter_outputs
for i, key_example in inputs_iterator:
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/iterable_dataset.py", line 1111, in iter_inputs
for key, example in iterator:
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/iterable_dataset.py", line 371, in __iter__
for key, pa_table in self.generate_tables_fn(**gen_kwags):
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/packaged_modules/json/json.py", line 99, in _generate_tables
for file_idx, file in enumerate(itertools.chain.from_iterable(files)):
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/utils/track.py", line 50, in __iter__
for x in self.generator(*self.args):
File "/miniconda3/envs/draft/lib/python3.11/site-packages/datasets/utils/file_utils.py", line 1378, in _iter_from_urlpaths
raise FileNotFoundError(urlpath)
FileNotFoundError: zstd://example_train_3889.jsonl::hf://datasets/cerebras/SlimPajama-627B@2d0accdd58c5d5511943ca1f5ff0e3eb5e293543/train/chunk9/example_train_3889.jsonl.zst
```
That final `raise` is at the bottom of the following snippet:
https://github.com/huggingface/datasets/blob/f693f4e93aabafa878470c80fd42ddb10ec550d6/src/datasets/utils/file_utils.py#L1354-L1379
So clearly, something choked up in `xisfile`.
### Steps to reproduce the bug
This happens when streaming a dataset and iterating over it. In my case, that iteration is done in Trainer's `inner_training_loop`, but this is not relevant to the iterator.
```python
File "/miniconda3/envs/draft/lib/python3.11/site-packages/accelerate/data_loader.py", line 835, in __iter__
next_batch, next_batch_info = self._fetch_batches(main_iterator)
```
### Expected behavior
This bug and the linked issue have one thing in common: *when streaming fails to retrieve an example, the entire program gives up and crashes*. As users, we cannot even protect ourselves from this: when we are iterating over a dataset, we can't make `datasets` skip over a bad example or wait a little longer to retry the iteration, because when a Python generator/iterator raises an error, it loses all its context.
In other words: if you have something that looks like `for b in a: for c in b: for d in c:`, errors in the innermost loop can only be caught by a `try ... except` in `c.__iter__()`. There should be such exception handling in `datasets` and it should have a **configurable exponential back-off**: first wait and retry after 1 minute, then 2 minutes, then 4 minutes, then 8 minutes, ... and after a given amount of retries, **skip the bad example**, and **only after** skipping a given amount of examples, give up and crash. This was requested in https://github.com/huggingface/datasets/issues/6843 too, since currently there is only linear backoff *and* it is clearly not applied to `xisfile`.
### Environment info
- `datasets` version: 3.3.2 *(the latest version)*
- Platform: Linux-4.18.0-513.24.1.el8_9.x86_64-x86_64-with-glibc2.28
- Python version: 3.11.7
- `huggingface_hub` version: 0.26.5
- PyArrow version: 15.0.0
- Pandas version: 2.2.0
- `fsspec` version: 2024.10.0
|
open
|
https://github.com/huggingface/datasets/issues/7440
| 2025-03-07T19:14:18 | 2025-07-22T08:15:44 | null |
{
"login": "bauwenst",
"id": 145220868,
"type": "User"
}
|
[] | false |
[] |
2,900,143,289 | 7,439 |
Fix multi gpu process example
|
to is not an inplace function.
But i am not sure about this code anyway, i think this is modifying the global variable `model` everytime the function is called? Which is on every batch? So it is juggling the same model on every gpu right? Isnt that very inefficient?
|
closed
|
https://github.com/huggingface/datasets/pull/7439
| 2025-03-06T11:29:19 | 2025-03-06T17:07:28 | 2025-03-06T17:06:38 |
{
"login": "SwayStar123",
"id": 46050679,
"type": "User"
}
|
[] | true |
[] |
2,899,209,484 | 7,438 |
Allow dataset row indexing with np.int types (#7423)
|
@lhoestq
Proposed fix for #7423. Added a couple simple tests as requested. I had some test failures related to Java and pyspark even when installing with dev but these don't seem to be related to the changes here and fail for me even on clean main.
The typeerror raised when using the wrong type is: "Wrong key type: '{key}' of type '{type(key)}'. Expected one of int, slice, range, str or Iterable." I think that is fine. But I could modify the int part to something more generic (although I'm not sure what) if wanted.
|
closed
|
https://github.com/huggingface/datasets/pull/7438
| 2025-03-06T03:10:43 | 2025-07-23T17:56:22 | 2025-07-23T16:44:42 |
{
"login": "DavidRConnell",
"id": 35470740,
"type": "User"
}
|
[] | true |
[] |
2,899,104,679 | 7,437 |
Use pyupgrade --py39-plus for remaining files
|
This work follows #7428. And "requires-python" is set in pyproject.toml
|
open
|
https://github.com/huggingface/datasets/pull/7437
| 2025-03-06T02:12:25 | 2025-07-18T04:04:08 | null |
{
"login": "cyyever",
"id": 17618148,
"type": "User"
}
|
[] | true |
[] |
2,898,385,725 | 7,436 |
chore: fix typos
| null |
closed
|
https://github.com/huggingface/datasets/pull/7436
| 2025-03-05T20:17:54 | 2025-04-28T14:00:09 | 2025-04-28T13:51:26 |
{
"login": "afuetterer",
"id": 35225576,
"type": "User"
}
|
[] | true |
[] |
2,895,536,956 | 7,435 |
Refactor `string_to_dict` to return `None` if there is no match instead of raising `ValueError`
|
Making this change, as encouraged here:
* https://github.com/huggingface/datasets/pull/7434#discussion_r1979933054
instead of having the pattern of using `try`-`except` to handle when there is no match, we can instead check if the return value is `None`; we can also assert that the return value should not be `None` if we know that should be true
|
closed
|
https://github.com/huggingface/datasets/pull/7435
| 2025-03-04T22:01:20 | 2025-03-12T16:52:00 | 2025-03-12T16:52:00 |
{
"login": "ringohoffman",
"id": 27844407,
"type": "User"
}
|
[] | true |
[] |
2,893,075,908 | 7,434 |
Refactor `Dataset.map` to reuse cache files mapped with different `num_proc`
|
Fixes #7433
This refactor unifies `num_proc is None or num_proc == 1` and `num_proc > 1`; instead of handling them completely separately where one uses a list of kwargs and shards and the other just uses a single set of kwargs and `self`, by wrapping the `num_proc == 1` case in a list and making the difference just whether or not you use a pool, you set up either case to be able to load each other's cache files just by changing `num_shards`; `num_proc == 1` can sequentially load the shards of a dataset mapped `num_shards > 1` and map any missing shards
Other than the structural refactor, the main contribution of this PR is `existing_cache_file_map`, which uses a regex of `cache_file_name` and `suffix_template` to find existing cache files, grouped by their `num_shards`; using this data structure, we can reset `num_shards` to an existing set of cache files, and load them accordingly
|
closed
|
https://github.com/huggingface/datasets/pull/7434
| 2025-03-04T06:12:37 | 2025-05-14T10:45:10 | 2025-05-12T15:14:08 |
{
"login": "ringohoffman",
"id": 27844407,
"type": "User"
}
|
[] | true |
[] |
2,890,240,400 | 7,433 |
`Dataset.map` ignores existing caches and remaps when ran with different `num_proc`
|
### Describe the bug
If you `map` a dataset and save it to a specific `cache_file_name` with a specific `num_proc`, and then call map again with that same existing `cache_file_name` but a different `num_proc`, the dataset will be re-mapped.
### Steps to reproduce the bug
1. Download a dataset
```python
import datasets
dataset = datasets.load_dataset("ylecun/mnist")
```
```
Generating train split: 100%|██████████| 60000/60000 [00:00<00:00, 116429.85 examples/s]
Generating test split: 100%|██████████| 10000/10000 [00:00<00:00, 103310.27 examples/s]
```
2. `map` and cache it with a specific `num_proc`
```python
cache_file_name="./cache/train.map"
dataset["train"].map(lambda x: x, cache_file_name=cache_file_name, num_proc=2)
```
```
Map (num_proc=2): 100%|██████████| 60000/60000 [00:01<00:00, 53764.03 examples/s]
```
3. `map` it with a different `num_proc` and the same `cache_file_name` as before
```python
dataset["train"].map(lambda x: x, cache_file_name=cache_file_name, num_proc=3)
```
```
Map (num_proc=3): 100%|██████████| 60000/60000 [00:00<00:00, 65377.12 examples/s]
```
### Expected behavior
If I specify an existing `cache_file_name`, I don't expect using a different `num_proc` than the one that was used to generate it to cause the dataset to have be be re-mapped.
### Environment info
```console
$ datasets-cli env
- `datasets` version: 3.3.2
- Platform: Linux-5.15.0-131-generic-x86_64-with-glibc2.35
- Python version: 3.10.16
- `huggingface_hub` version: 0.29.1
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
```
|
closed
|
https://github.com/huggingface/datasets/issues/7433
| 2025-03-03T05:51:26 | 2025-05-12T15:14:09 | 2025-05-12T15:14:09 |
{
"login": "ringohoffman",
"id": 27844407,
"type": "User"
}
|
[] | false |
[] |
2,887,717,289 | 7,432 |
Fix type annotation
| null |
closed
|
https://github.com/huggingface/datasets/pull/7432
| 2025-02-28T17:28:20 | 2025-03-04T15:53:03 | 2025-03-04T15:53:03 |
{
"login": "NeilGirdhar",
"id": 730137,
"type": "User"
}
|
[] | true |
[] |
2,887,244,074 | 7,431 |
Issues with large Datasets
|
### Describe the bug
If the coco annotation file is too large the dataset will not be able to load it, not entirely sure were the issue is but I am guessing it is due to the code trying to load it all as one line into a dataframe. This was for object detections.
My current work around is the following code but would be nice to be able to do it without worrying about it also probably there is a better way of doing it:
`
dataset_dict = json.load(open("./local_data/annotations/train.json"))
df = pd.DataFrame(columns=['images', 'annotations', 'categories'])
df = df._append({'images': dataset_dict['images'], 'annotations': dataset_dict['annotations'], 'categories': dataset_dict['categories']}, ignore_index=True)
train=Dataset.from_pandas(df)
dataset_dict = json.load(open("./local_data/annotations/validation.json"))
df = pd.DataFrame(columns=['images', 'annotations', 'categories'])
df = df._append({'images': dataset_dict['images'], 'annotations': dataset_dict['annotations'],
'categories': dataset_dict['categories']}, ignore_index=True)
val = Dataset.from_pandas(df)
dataset_dict = json.load(open("./local_data/annotations/test.json"))
df = pd.DataFrame(columns=['images', 'annotations', 'categories'])
df = df._append({'images': dataset_dict['images'], 'annotations': dataset_dict['annotations'],
'categories': dataset_dict['categories']}, ignore_index=True)
test = Dataset.from_pandas(df)
dataset = DatasetDict({'train': train, 'validation': val, 'test': test})
`
### Steps to reproduce the bug
1) step up directory in and have the json files in coco format
-local_data
|-images
|---1.jpg
|---2.jpg
|---....
|---n.jpg
|-annotations
|---test.json
|---train.json
|---validation.json
2) try to load local_data into a dataset if the file is larger than about 300kb it will cause an error.
### Expected behavior
That it loads the jsons preferably in the same format as it has done with a smaller size.
### Environment info
- `datasets` version: 3.3.3.dev0
- Platform: Linux-6.11.0-17-generic-x86_64-with-glibc2.39
- Python version: 3.12.3
- `huggingface_hub` version: 0.29.0
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
|
open
|
https://github.com/huggingface/datasets/issues/7431
| 2025-02-28T14:05:22 | 2025-03-04T15:02:26 | null |
{
"login": "nikitabelooussovbtis",
"id": 106806889,
"type": "User"
}
|
[] | false |
[] |
2,886,922,573 | 7,430 |
Error in code "Time to slice and dice" from course "NLP Course"
|
### Describe the bug
When we execute code
```
frequencies = (
train_df["condition"]
.value_counts()
.to_frame()
.reset_index()
.rename(columns={"index": "condition", "condition": "frequency"})
)
frequencies.head()
```
answer should be like this
condition | frequency
birth control | 27655
depression | 8023
acne | 5209
anxiety | 4991
pain | 4744
but he is different
frequency | count
birth control | 27655
depression | 8023
acne | 5209
anxiety | 4991
pain | 4744
this is not correct, correct code
```
frequencies = (
train_df["condition"]
.value_counts()
.to_frame()
.reset_index()
.rename(columns={"index": "condition", "count": "frequency"})
)
````
### Steps to reproduce the bug
```
frequencies = (
train_df["condition"]
.value_counts()
.to_frame()
.reset_index()
.rename(columns={"index": "condition", "condition": "frequency"})
)
frequencies.head()
```
### Expected behavior
condition | frequency
birth control | 27655
depression | 8023
acne | 5209
anxiety | 4991
pain | 4744
### Environment info
Google Colab
|
closed
|
https://github.com/huggingface/datasets/issues/7430
| 2025-02-28T11:36:10 | 2025-03-05T11:32:47 | 2025-03-03T17:52:15 |
{
"login": "Yurkmez",
"id": 122965300,
"type": "User"
}
|
[] | false |
[] |
2,886,806,513 | 7,429 |
Improved type annotation
|
I've refined several type annotations throughout the codebase to align with current best practices and enhance overall clarity. Given the complexity of the code, there may still be areas that need further attention. I welcome any feedback or suggestions to make these improvements even better.
- Fixes #7202
|
open
|
https://github.com/huggingface/datasets/pull/7429
| 2025-02-28T10:39:10 | 2025-05-15T12:27:17 | null |
{
"login": "saiden89",
"id": 45285915,
"type": "User"
}
|
[] | true |
[] |
2,886,111,651 | 7,428 |
Use pyupgrade --py39-plus
| null |
closed
|
https://github.com/huggingface/datasets/pull/7428
| 2025-02-28T03:39:44 | 2025-03-22T00:51:20 | 2025-03-05T15:04:16 |
{
"login": "cyyever",
"id": 17618148,
"type": "User"
}
|
[] | true |
[] |
2,886,032,571 | 7,427 |
Error splitting the input into NAL units.
|
### Describe the bug
I am trying to finetune qwen2.5-vl on 16 * 80G GPUS, and I use `LLaMA-Factory` and set `preprocessing_num_workers=16`. However, I met the following error and the program seem to got crush. It seems that the error come from `datasets` library
The error logging is like following:
```text
Converting format of dataset (num_proc=16): 100%|█████████▉| 19265/19267 [11:44<00:00, 5.88 examples/s]
Converting format of dataset (num_proc=16): 100%|█████████▉| 19266/19267 [11:44<00:00, 5.02 examples/s]
Converting format of dataset (num_proc=16): 100%|██████████| 19267/19267 [11:44<00:00, 5.44 examples/s]
Converting format of dataset (num_proc=16): 100%|██████████| 19267/19267 [11:44<00:00, 27.34 examples/s]
Running tokenizer on dataset (num_proc=16): 0%| | 0/19267 [00:00<?, ? examples/s]
Invalid NAL unit size (45405 > 35540).
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (7131 > 3225).
missing picture in access unit with size 54860
Invalid NAL unit size (48042 > 33645).
missing picture in access unit with size 3229
missing picture in access unit with size 33649
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (48042 > 33645).
Error splitting the input into NAL units.
missing picture in access unit with size 35544
Invalid NAL unit size (45405 > 35540).
Error splitting the input into NAL units.
Error splitting the input into NAL units.
Invalid NAL unit size (8187 > 7069).
missing picture in access unit with size 7073
Invalid NAL unit size (8187 > 7069).
Error splitting the input into NAL units.
Invalid NAL unit size (7131 > 3225).
Error splitting the input into NAL units.
Invalid NAL unit size (14013 > 5998).
missing picture in access unit with size 6002
Invalid NAL unit size (14013 > 5998).
Error splitting the input into NAL units.
Invalid NAL unit size (17173 > 7231).
missing picture in access unit with size 7235
Invalid NAL unit size (17173 > 7231).
Error splitting the input into NAL units.
Invalid NAL unit size (16964 > 6055).
missing picture in access unit with size 6059
Invalid NAL unit size (16964 > 6055).
Exception in thread Thread-9 (accepter)Error splitting the input into NAL units.
:
Traceback (most recent call last):
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 1016, in _bootstrap_inner
Running tokenizer on dataset (num_proc=16): 0%| | 0/19267 [13:22<?, ? examples/s] self.run()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 953, in run
Invalid NAL unit size (7032 > 2927).
missing picture in access unit with size 2931
self._target(*self._args, **self._kwargs)
File "/opt/conda/envs/python3.10.13/lib/python3.10/site-packages/multiprocess/managers.py", line 194, in accepter
Invalid NAL unit size (7032 > 2927).
Error splitting the input into NAL units.
t.start()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 935, in start
Invalid NAL unit size (28973 > 6121).
missing picture in access unit with size 6125
_start_new_thread(self._bootstrap, ())Invalid NAL unit size (28973 > 6121).
RuntimeError: can't start new threadError splitting the input into NAL units.
Invalid NAL unit size (4411 > 296).
missing picture in access unit with size 300
Invalid NAL unit size (4411 > 296).
Error splitting the input into NAL units.
Invalid NAL unit size (14414 > 1471).
missing picture in access unit with size 1475
Invalid NAL unit size (14414 > 1471).
Error splitting the input into NAL units.
Invalid NAL unit size (5283 > 1792).
missing picture in access unit with size 1796
Invalid NAL unit size (5283 > 1792).
Error splitting the input into NAL units.
Invalid NAL unit size (79147 > 10042).
missing picture in access unit with size 10046
Invalid NAL unit size (79147 > 10042).
Error splitting the input into NAL units.
Invalid NAL unit size (45405 > 35540).
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (7131 > 3225).
missing picture in access unit with size 54860
Invalid NAL unit size (48042 > 33645).
missing picture in access unit with size 3229
missing picture in access unit with size 33649
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (48042 > 33645).
Error splitting the input into NAL units.
missing picture in access unit with size 35544
Invalid NAL unit size (45405 > 35540).
Error splitting the input into NAL units.
Error splitting the input into NAL units.
Invalid NAL unit size (8187 > 7069).
missing picture in access unit with size 7073
Invalid NAL unit size (8187 > 7069).
Error splitting the input into NAL units.
Invalid NAL unit size (7131 > 3225).
Error splitting the input into NAL units.
Invalid NAL unit size (14013 > 5998).
missing picture in access unit with size 6002
Invalid NAL unit size (14013 > 5998).
Error splitting the input into NAL units.
Invalid NAL unit size (17173 > 7231).
missing picture in access unit with size 7235
Invalid NAL unit size (17173 > 7231).
Error splitting the input into NAL units.
Invalid NAL unit size (16964 > 6055).
missing picture in access unit with size 6059
Invalid NAL unit size (16964 > 6055).
Exception in thread Thread-9 (accepter)Error splitting the input into NAL units.
:
Traceback (most recent call last):
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 1016, in _bootstrap_inner
Running tokenizer on dataset (num_proc=16): 0%| | 0/19267 [13:22<?, ? examples/s] self.run()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 953, in run
Invalid NAL unit size (7032 > 2927).
missing picture in access unit with size 2931
self._target(*self._args, **self._kwargs)
File "/opt/conda/envs/python3.10.13/lib/python3.10/site-packages/multiprocess/managers.py", line 194, in accepter
Invalid NAL unit size (7032 > 2927).
Error splitting the input into NAL units.
t.start()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 935, in start
Invalid NAL unit size (28973 > 6121).
missing picture in access unit with size 6125
_start_new_thread(self._bootstrap, ())Invalid NAL unit size (28973 > 6121).
RuntimeError: can't start new threadError splitting the input into NAL units.
Invalid NAL unit size (4411 > 296).
missing picture in access unit with size 300
Invalid NAL unit size (4411 > 296).
Error splitting the input into NAL units.
Invalid NAL unit size (14414 > 1471).
missing picture in access unit with size 1475
Invalid NAL unit size (14414 > 1471).
Error splitting the input into NAL units.
Invalid NAL unit size (5283 > 1792).
missing picture in access unit with size 1796
Invalid NAL unit size (5283 > 1792).
Error splitting the input into NAL units.
Invalid NAL unit size (79147 > 10042).
missing picture in access unit with size 10046
Invalid NAL unit size (79147 > 10042).
Error splitting the input into NAL units.
Invalid NAL unit size (45405 > 35540).
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (7131 > 3225).
missing picture in access unit with size 54860
Invalid NAL unit size (48042 > 33645).
missing picture in access unit with size 3229
missing picture in access unit with size 33649
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (48042 > 33645).
Error splitting the input into NAL units.
missing picture in access unit with size 35544
Invalid NAL unit size (45405 > 35540).
Error splitting the input into NAL units.
Error splitting the input into NAL units.
Invalid NAL unit size (8187 > 7069).
missing picture in access unit with size 7073
Invalid NAL unit size (8187 > 7069).
Error splitting the input into NAL units.
Invalid NAL unit size (7131 > 3225).
Error splitting the input into NAL units.
Invalid NAL unit size (14013 > 5998).
missing picture in access unit with size 6002
Invalid NAL unit size (14013 > 5998).
Error splitting the input into NAL units.
Invalid NAL unit size (17173 > 7231).
missing picture in access unit with size 7235
Invalid NAL unit size (17173 > 7231).
Error splitting the input into NAL units.
Invalid NAL unit size (16964 > 6055).
missing picture in access unit with size 6059
Invalid NAL unit size (16964 > 6055).
Exception in thread Thread-9 (accepter)Error splitting the input into NAL units.
:
Traceback (most recent call last):
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 1016, in _bootstrap_inner
Running tokenizer on dataset (num_proc=16): 0%| | 0/19267 [13:22<?, ? examples/s] self.run()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 953, in run
Invalid NAL unit size (7032 > 2927).
missing picture in access unit with size 2931
self._target(*self._args, **self._kwargs)
File "/opt/conda/envs/python3.10.13/lib/python3.10/site-packages/multiprocess/managers.py", line 194, in accepter
Invalid NAL unit size (7032 > 2927).
Error splitting the input into NAL units.
t.start()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 935, in start
Invalid NAL unit size (28973 > 6121).
missing picture in access unit with size 6125
_start_new_thread(self._bootstrap, ())Invalid NAL unit size (28973 > 6121).
RuntimeError: can't start new threadError splitting the input into NAL units.
Invalid NAL unit size (4411 > 296).
missing picture in access unit with size 300
Invalid NAL unit size (4411 > 296).
Error splitting the input into NAL units.
Invalid NAL unit size (14414 > 1471).
missing picture in access unit with size 1475
Invalid NAL unit size (14414 > 1471).
Error splitting the input into NAL units.
Invalid NAL unit size (5283 > 1792).
missing picture in access unit with size 1796
Invalid NAL unit size (5283 > 1792).
Error splitting the input into NAL units.
Invalid NAL unit size (79147 > 10042).
missing picture in access unit with size 10046
Invalid NAL unit size (79147 > 10042).
Error splitting the input into NAL units.
Invalid NAL unit size (45405 > 35540).
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (7131 > 3225).
missing picture in access unit with size 54860
Invalid NAL unit size (48042 > 33645).
missing picture in access unit with size 3229
missing picture in access unit with size 33649
Invalid NAL unit size (86720 > 54856).
Invalid NAL unit size (48042 > 33645).
Error splitting the input into NAL units.
missing picture in access unit with size 35544
Invalid NAL unit size (45405 > 35540).
Error splitting the input into NAL units.
Error splitting the input into NAL units.
Invalid NAL unit size (8187 > 7069).
missing picture in access unit with size 7073
Invalid NAL unit size (8187 > 7069).
Error splitting the input into NAL units.
Invalid NAL unit size (7131 > 3225).
Error splitting the input into NAL units.
Invalid NAL unit size (14013 > 5998).
missing picture in access unit with size 6002
Invalid NAL unit size (14013 > 5998).
Error splitting the input into NAL units.
Invalid NAL unit size (17173 > 7231).
missing picture in access unit with size 7235
Invalid NAL unit size (17173 > 7231).
Error splitting the input into NAL units.
Invalid NAL unit size (16964 > 6055).
missing picture in access unit with size 6059
Invalid NAL unit size (16964 > 6055).
Exception in thread Thread-9 (accepter)Error splitting the input into NAL units.
:
Traceback (most recent call last):
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 1016, in _bootstrap_inner
Running tokenizer on dataset (num_proc=16): 0%| | 0/19267 [13:22<?, ? examples/s] self.run()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 953, in run
Invalid NAL unit size (7032 > 2927).
missing picture in access unit with size 2931
self._target(*self._args, **self._kwargs)
File "/opt/conda/envs/python3.10.13/lib/python3.10/site-packages/multiprocess/managers.py", line 194, in accepter
Invalid NAL unit size (7032 > 2927).
Error splitting the input into NAL units.
t.start()
File "/opt/conda/envs/python3.10.13/lib/python3.10/threading.py", line 935, in start
Invalid NAL unit size (28973 > 6121).
missing picture in access unit with size 6125
_start_new_thread(self._bootstrap, ())Invalid NAL unit size (28973 > 6121).
RuntimeError: can't start new threadError splitting the input into NAL units.
Invalid NAL unit size (4411 > 296).
missing picture in access unit with size 300
Invalid NAL unit size (4411 > 296).
Error splitting the input into NAL units.
Invalid NAL unit size (14414 > 1471).
missing picture in access unit with size 1475
Invalid NAL unit size (14414 > 1471).
Error splitting the input into NAL units.
Invalid NAL unit size (5283 > 1792).
missing picture in access unit with size 1796
Invalid NAL unit size (5283 > 1792).
Error splitting the input into NAL units.
Invalid NAL unit size (79147 > 10042).
missing picture in access unit with size 10046
Invalid NAL unit size (79147 > 10042).
Error splitting the input into NAL units.
```
### Others
_No response_
### Steps to reproduce the bug
None
### Expected behavior
excpect to run successfully
### Environment info
```
transformers==4.49.0
datasets==3.2.0
accelerate==1.2.1
peft==0.12.0
trl==0.9.6
tokenizers==0.21.0
gradio>=4.38.0,<=5.18.0
pandas>=2.0.0
scipy
einops
sentencepiece
tiktoken
protobuf
uvicorn
pydantic
fastapi
sse-starlette
matplotlib>=3.7.0
fire
packaging
pyyaml
numpy<2.0.0
av
librosa
tyro<0.9.0
openlm-hub
qwen-vl-utils
```
|
open
|
https://github.com/huggingface/datasets/issues/7427
| 2025-02-28T02:30:15 | 2025-03-04T01:40:28 | null |
{
"login": "MengHao666",
"id": 47114466,
"type": "User"
}
|
[] | false |
[] |
2,883,754,507 | 7,426 |
fix: None default with bool type on load creates typing error
|
Hello!
Pyright flags any use of `load_dataset` as an error, because the default for `trust_remote_code` is `None`, but the function is typed as `bool`, not `Optional[bool]`. I changed the type and docstrings to reflect this, but no other code was touched.
|
closed
|
https://github.com/huggingface/datasets/pull/7426
| 2025-02-27T08:11:36 | 2025-03-04T15:53:40 | 2025-03-04T15:53:40 |
{
"login": "stephantul",
"id": 8882233,
"type": "User"
}
|
[] | true |
[] |
2,883,684,686 | 7,425 |
load_dataset("livecodebench/code_generation_lite", version_tag="release_v2") TypeError: 'NoneType' object is not callable
|
### Describe the bug
from datasets import load_dataset
lcb_codegen = load_dataset("livecodebench/code_generation_lite", version_tag="release_v2")
or
configs = get_dataset_config_names("livecodebench/code_generation_lite", trust_remote_code=True)
both error:
Traceback (most recent call last):
File "", line 1, in
File "/workspace/miniconda/envs/grpo/lib/python3.10/site-packages/datasets/load.py", line 2131, in load_dataset
builder_instance = load_dataset_builder(
File "/workspace/miniconda/envs/grpo/lib/python3.10/site-packages/datasets/load.py", line 1888, in load_dataset_builder
builder_instance: DatasetBuilder = builder_cls(
TypeError: 'NoneType' object is not callable
### Steps to reproduce the bug
from datasets import get_dataset_config_names
configs = get_dataset_config_names("livecodebench/code_generation_lite", trust_remote_code=True)
OR
lcb_codegen = load_dataset("livecodebench/code_generation_lite", version_tag="release_v2")
### Expected behavior
load datasets livecodebench/code_generation_lite
### Environment info
import datasets
version '3.3.2'
|
open
|
https://github.com/huggingface/datasets/issues/7425
| 2025-02-27T07:36:02 | 2025-03-27T05:05:33 | null |
{
"login": "dshwei",
"id": 42167236,
"type": "User"
}
|
[] | false |
[] |
2,882,663,621 | 7,424 |
Faster folder based builder + parquet support + allow repeated media + use torchvideo
|
This will be useful for LeRobotDataset (robotics datasets for [lerobot](https://github.com/huggingface/lerobot) based on videos)
Impacted builders:
- ImageFolder
- AudioFolder
- VideoFolder
Improvements:
- faster to stream (got a 5x speed up on an image dataset)
- improved RAM usage
- support for metadata.parquet
- allow to link to an image/audio/video multiple times
- support for pyarrow filters (mostly efficient for parquet)
- link to files using fields names `*_file_name` (in addition to the already existing `file_name`)
- this allows to have multiple image/audio/video per row
- there is also `file_names` and `*_file_names` for lists of image/audio/video
Changes:
- the builders iterate on the metadata files instead of the media files
- the builders iterate on chunks of metadata instead of loading them in RAM completely
- metadata files are no longer handled separately in `data_files`
- added the `filters` argument to pass to `load_dataset`
- either as an [Expression](https://arrow.apache.org/docs/python/generated/pyarrow.dataset.Expression.html)
- or as tuples like `filters=[('event_name', '=', 'SomeEvent')]`
- small breaking change: you can't add labels to a dataset with`drop_labels=False` if it has a metadata file
- small breaking change: you can't use one metadata file for multiple splits anymore
Example: `lhoestq/pusht-videofolder` is a video dataset with metadata.parquet where multiple rows can point to the same video
```python
In [1]: from datasets import load_dataset
In [2]: load_dataset("lhoestq/pusht-videofolder")
Resolving data files: 100%|██████████████████████████████| 207/207 [00:00<00:00, 1087.32it/s]
Out[2]:
DatasetDict({
train: Dataset({
features: ['video', 'observation.state', 'action', 'episode_index', 'frame_index', 'timestamp', 'next.reward', 'next.done', 'next.success', 'index', 'task_index'],
num_rows: 25650
})
})
In [3]: load_dataset("lhoestq/pusht-videofolder", filters=[("next.reward", ">", 0.5)])
Resolving data files: 100%|██████████████████████████████| 207/207 [00:01<00:00, 183.03it/s]
Out[3]:
DatasetDict({
train: Dataset({
features: ['video', 'observation.state', 'action', 'episode_index', 'frame_index', 'timestamp', 'next.reward', 'next.done', 'next.success', 'index', 'task_index'],
num_rows: 5773
})
})
```
Additional change for VideoFolder:
- decord can't be installed in many setups, I switched the backend to torchvision instead
- I also added streaming capability from HF (you can get video frames without downloading the full video from HF)
Example: load a robotics dataset
```python
In [1]: from datasets import load_dataset
ds
In [2]: ds = load_dataset("lhoestq/pusht-videofolder")
Resolving data files: 100%|██████████████████████████████| 207/207 [00:00<00:00, 624.81it/s]
In [3]: ds["train"][0]
Out[3]:
{'video': <torchvision.io.video_reader.VideoReader at 0x1145dc290>,
'observation.state': [222.0, 97.0],
'action': [233.0, 71.0],
'episode_index': 0,
'frame_index': 0,
'timestamp': 0.0,
'next.reward': 0.19029748439788818,
'next.done': False,
'next.success': False,
'index': 0,
'task_index': 0}
```
Example: stream frames without downloading full videos
```python
In [1]: from datasets import load_dataset
In [2]: ds = load_dataset("BrianGuo/Tennis_Data", streaming=True)
In [3]: example = next(iter(ds["train"]))
In [4]: video = example["video"]
In [5]: video.get_metadata()
Out[5]:
{'audio': {'framerate': [44100.0], 'duration': [2027.35]},
'video': {'fps': [59.00002712894387], 'duration': [2027.355]}}
In [6]: video.seek(1800, keyframes_only=True) # 30min
Out[6]: <torchvision.io.video_reader.VideoReader at 0x148d4d010>
In [7]: next(video)
Out[7]:
{'data': tensor([[[ 76, 77, 79, ..., 41, 39, 38],
[ 76, 77, 79, ..., 40, 39, 35],
[ 76, 77, 79, ..., 34, 30, 26],
...,
[127, 127, 127, ..., 125, 125, 125],
[125, 126, 126, ..., 125, 125, 125],
[122, 124, 126, ..., 125, 125, 125]]], dtype=torch.uint8),
'pts': 1800.0}
```
TODO:
- [x] docs
- [x] fix tests
|
closed
|
https://github.com/huggingface/datasets/pull/7424
| 2025-02-26T19:55:18 | 2025-03-05T18:51:00 | 2025-03-05T17:41:23 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,879,271,409 | 7,423 |
Row indexing a dataset with numpy integers
|
### Feature request
Allow indexing datasets with a scalar numpy integer type.
### Motivation
Indexing a dataset with a scalar numpy.int* object raises a TypeError. This is due to the test in `datasets/formatting/formatting.py:key_to_query_type`
``` python
def key_to_query_type(key: Union[int, slice, range, str, Iterable]) -> str:
if isinstance(key, int):
return "row"
elif isinstance(key, str):
return "column"
elif isinstance(key, (slice, range, Iterable)):
return "batch"
_raise_bad_key_type(key)
```
In the row case, it checks if key is an int, which returns false when key is integer like but not a builtin python integer type. This is counterintuitive because a numpy array of np.int64s can be used for the batch case.
For example:
``` python
import numpy as np
import datasets
dataset = datasets.Dataset.from_dict({"a": [1, 2, 3, 4], "b": [5, 6, 7, 8]})
# Regular indexing
dataset[0]
dataset[:2]
# Indexing with numpy data types (expect same results)
idx = np.asarray([0, 1])
dataset[idx] # Succeeds when using an array of np.int64 values
dataset[idx[0]] # Fails with TypeError when using scalar np.int64
```
For the user, this can be solved by wrapping `idx[0]` in `int` but the test could also be changed in `key_to_query_type` to accept a less strict definition of int.
``` diff
+import numbers
+
def key_to_query_type(key: Union[int, slice, range, str, Iterable]) -> str:
+ if isinstance(key, numbers.Integral):
- if isinstance(key, int):
return "row"
elif isinstance(key, str):
return "column"
elif isinstance(key, (slice, range, Iterable)):
return "batch"
_raise_bad_key_type(key)
```
Looking at how others do it, pandas has an `is_integer` definition that it checks which uses `is_integer_object` defined in `pandas/_libs/utils.pxd`:
``` cython
cdef inline bint is_integer_object(object obj) noexcept:
"""
Cython equivalent of
`isinstance(val, (int, np.integer)) and not isinstance(val, (bool, np.timedelta64))`
Parameters
----------
val : object
Returns
-------
is_integer : bool
Notes
-----
This counts np.timedelta64 objects as integers.
"""
return (not PyBool_Check(obj) and isinstance(obj, (int, cnp.integer))
and not is_timedelta64_object(obj))
```
This would be less flexible as it explicitly checks for numpy integer, but worth noting that they had the need to ensure the key is not a bool.
### Your contribution
I can submit a pull request with the above changes after checking that indexing succeeds with the numpy integer type. Or if there is a different integer check that would be preferred I could add that.
If there is a reason not to want this behavior that is fine too.
|
open
|
https://github.com/huggingface/datasets/issues/7423
| 2025-02-25T18:44:45 | 2025-03-03T17:55:24 | null |
{
"login": "DavidRConnell",
"id": 35470740,
"type": "User"
}
|
[
{
"name": "enhancement",
"color": "a2eeef"
}
] | false |
[] |
2,878,369,052 | 7,421 |
DVC integration broken
|
### Describe the bug
The DVC integration seems to be broken.
Followed this guide: https://dvc.org/doc/user-guide/integrations/huggingface
### Steps to reproduce the bug
#### Script to reproduce
~~~python
from datasets import load_dataset
dataset = load_dataset(
"csv",
data_files="dvc://workshop/satellite-data/jan_train.csv",
storage_options={"url": "https://github.com/iterative/dataset-registry.git"},
)
print(dataset)
~~~
#### Error log
~~~
Traceback (most recent call last):
File "C:\tmp\test\load.py", line 3, in <module>
dataset = load_dataset(
^^^^^^^^^^^^^
File "C:\tmp\test\.venv\Lib\site-packages\datasets\load.py", line 2151, in load_dataset
builder_instance.download_and_prepare(
File "C:\tmp\test\.venv\Lib\site-packages\datasets\builder.py", line 808, in download_and_prepare
fs, output_dir = url_to_fs(output_dir, **(storage_options or {}))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: url_to_fs() got multiple values for argument 'url'
~~~
### Expected behavior
Integration would work and the indicated file is downloaded and opened.
### Environment info
#### Python version
~~~
python --version
Python 3.11.10
~~~
#### Venv (pip install datasets dvc):
~~~
Package Version
---------------------- -----------
aiohappyeyeballs 2.4.6
aiohttp 3.11.13
aiohttp-retry 2.9.1
aiosignal 1.3.2
amqp 5.3.1
annotated-types 0.7.0
antlr4-python3-runtime 4.9.3
appdirs 1.4.4
asyncssh 2.20.0
atpublic 5.1
attrs 25.1.0
billiard 4.2.1
celery 5.4.0
certifi 2025.1.31
cffi 1.17.1
charset-normalizer 3.4.1
click 8.1.8
click-didyoumean 0.3.1
click-plugins 1.1.1
click-repl 0.3.0
colorama 0.4.6
configobj 5.0.9
cryptography 44.0.1
datasets 3.3.2
dictdiffer 0.9.0
dill 0.3.8
diskcache 5.6.3
distro 1.9.0
dpath 2.2.0
dulwich 0.22.7
dvc 3.59.1
dvc-data 3.16.9
dvc-http 2.32.0
dvc-objects 5.1.0
dvc-render 1.0.2
dvc-studio-client 0.21.0
dvc-task 0.40.2
entrypoints 0.4
filelock 3.17.0
flatten-dict 0.4.2
flufl-lock 8.1.0
frozenlist 1.5.0
fsspec 2024.12.0
funcy 2.0
gitdb 4.0.12
gitpython 3.1.44
grandalf 0.8
gto 1.7.2
huggingface-hub 0.29.1
hydra-core 1.3.2
idna 3.10
iterative-telemetry 0.0.10
kombu 5.4.2
markdown-it-py 3.0.0
mdurl 0.1.2
multidict 6.1.0
multiprocess 0.70.16
networkx 3.4.2
numpy 2.2.3
omegaconf 2.3.0
orjson 3.10.15
packaging 24.2
pandas 2.2.3
pathspec 0.12.1
platformdirs 4.3.6
prompt-toolkit 3.0.50
propcache 0.3.0
psutil 7.0.0
pyarrow 19.0.1
pycparser 2.22
pydantic 2.10.6
pydantic-core 2.27.2
pydot 3.0.4
pygit2 1.17.0
pygments 2.19.1
pygtrie 2.5.0
pyparsing 3.2.1
python-dateutil 2.9.0.post0
pytz 2025.1
pywin32 308
pyyaml 6.0.2
requests 2.32.3
rich 13.9.4
ruamel-yaml 0.18.10
ruamel-yaml-clib 0.2.12
scmrepo 3.3.10
semver 3.0.4
setuptools 75.8.0
shellingham 1.5.4
shortuuid 1.0.13
shtab 1.7.1
six 1.17.0
smmap 5.0.2
sqltrie 0.11.2
tabulate 0.9.0
tomlkit 0.13.2
tqdm 4.67.1
typer 0.15.1
typing-extensions 4.12.2
tzdata 2025.1
urllib3 2.3.0
vine 5.1.0
voluptuous 0.15.2
wcwidth 0.2.13
xxhash 3.5.0
yarl 1.18.3
zc-lockfile 3.0.post1
~~~
|
open
|
https://github.com/huggingface/datasets/issues/7421
| 2025-02-25T13:14:31 | 2025-03-03T17:42:02 | null |
{
"login": "maxstrobel",
"id": 34747372,
"type": "User"
}
|
[] | false |
[] |
2,876,281,928 | 7,420 |
better correspondence between cached and saved datasets created using from_generator
|
### Feature request
At the moment `.from_generator` can only create a dataset that lives in the cache. The cached dataset cannot be loaded with `load_from_disk` because the cache folder is missing `state.json`. So the only way to convert this cached dataset to a regular is to use `save_to_disk` which needs to create a copy of the cached dataset. For large datasets this can end up wasting a lot of space. In my case the saving operation failed so I am stuck with a large cached dataset and no clear way to convert to a `Dataset` that I can use. The requested feature is to provide a way to be able to load a cached dataset using `.load_from_disk`. Alternatively `.from_generator` can create the dataset at a specified location so that it can be loaded from there with `.load_from_disk`.
### Motivation
I have the following workflow which has exposed some awkwardness about the Datasets saving/caching.
1. I created a cached dataset using `.from_generator` which was cached in a folder. This dataset is rather large (~600GB) with many shards.
2. I tried to save this dataset using `.save_to_disk` to another location so that I can use later as a `Dataset`. This essentially creates another copy (for a total of 1.2TB!) of what is already in the cache... In my case the saving operation keeps dying for some reason and I am stuck with a cached dataset and no copy.
3. Now I am trying to "save" the existing cached dataset but it is not clear how to access the cached files after `.from_generator` has finished e.g. from a different process. I should not be even looking at the cache but I really do not want to waste another 2hr to generate the set so that if fails agains (I already did this couple of times).
- I tried `.load_from_disk` but it does not work with cached files and complains that this is not a `Dataset` (!).
- I looked at `.from_file` which takes one file but the cached file has many (shards) so I am not sure how to make this work.
- I tried `.load_dataset` but this seems to either try to "download" a copy (of a file which is already in the local file system!) which I will then need to save or I need to use `streaming=False` to create an `IterableDataset `which then I need to convert (using the cache) to `Dataset` so that I can save it. With both options I will end up with 3 copies of the same dataset for a total of ~2TB! I am hoping here is another way to do this...
Maybe I am missing something here: I looked at docs and forums but no luck. I have a bunch of arrow files cached by `Dataset.from_generator` and no clean way to make them into a `Dataset` that I can use.
This all could be so much easer if `load_from_disk` can recognize the cached files and produce a `Dataset`: after the cache is created I would not have to "save" it again and I can just load it when I need. At the moment `load_from_disk` needs `state.json` which is lacking in the cache folder. So perhaps `.from_generator` could be made to "finalize" (e.g. create `state.json`) the dataset once it is done so that it can be loaded easily. Or provide `.from_generator` with a `save_to_dir` parameter in addition to `cache_dir` which can be used for the whole process including creating the `state.json` at the end.
As a proof of concept I just created `state.json` by hand and `load_from_disk` worked using the cache! So it seems to be the missing piece here.
### Your contribution
Time permitting I can look into `.from_generator` to see if adding `state.json` is feasible.
|
open
|
https://github.com/huggingface/datasets/issues/7420
| 2025-02-24T22:14:37 | 2025-02-26T03:10:22 | null |
{
"login": "vttrifonov",
"id": 12157034,
"type": "User"
}
|
[
{
"name": "enhancement",
"color": "a2eeef"
}
] | false |
[] |
2,875,635,320 | 7,419 |
Import order crashes script execution
|
### Describe the bug
Hello,
I'm trying to convert an HF dataset into a TFRecord so I'm importing `tensorflow` and `datasets` to do so.
Depending in what order I'm importing those librairies, my code hangs forever and is unkillable (CTRL+C doesn't work, I need to kill my shell entirely).
Thank you for your help
🙏
### Steps to reproduce the bug
If you run the following script, this will hang forever :
```python
import tensorflow as tf
import datasets
dataset = datasets.load_dataset("imagenet-1k", split="validation", streaming=True)
print(next(iter(dataset)))
```
however running the following will work fine (I just changed the order of the imports) :
```python
import datasets
import tensorflow as tf
dataset = datasets.load_dataset("imagenet-1k", split="validation", streaming=True)
print(next(iter(dataset)))
```
### Expected behavior
I'm expecting the script to reach the end and my case print the content of the first item in the dataset
```
{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=408x500 at 0x70C646A03110>, 'label': 91}
```
### Environment info
```
$ datasets-cli env
- `datasets` version: 3.3.2
- Platform: Linux-6.8.0-1017-aws-x86_64-with-glibc2.35
- Python version: 3.11.7
- `huggingface_hub` version: 0.29.1
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
```
I'm also using `tensorflow==2.18.0`.
|
open
|
https://github.com/huggingface/datasets/issues/7419
| 2025-02-24T17:03:43 | 2025-02-24T17:03:43 | null |
{
"login": "DamienMatias",
"id": 23298479,
"type": "User"
}
|
[] | false |
[] |
2,868,701,471 | 7,418 |
pyarrow.lib.arrowinvalid: cannot mix list and non-list, non-null values with map function
|
### Describe the bug
Encounter pyarrow.lib.arrowinvalid error with map function in some example when loading the dataset
### Steps to reproduce the bug
```
from datasets import load_dataset
from PIL import Image, PngImagePlugin
dataset = load_dataset("leonardPKU/GEOQA_R1V_Train_8K")
system_prompt="You are a helpful AI Assistant"
def make_conversation(example):
prompt = []
prompt.append({"role": "system", "content": system_prompt})
prompt.append(
{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": example["problem"]},
]
}
)
return {"prompt": prompt}
def check_data_types(example):
for key, value in example.items():
if key == 'image':
if not isinstance(value, PngImagePlugin.PngImageFile):
print(value)
if key == "problem" or key == "solution":
if not isinstance(value, str):
print(value)
return example
dataset = dataset.map(check_data_types)
dataset = dataset.map(make_conversation)
```
### Expected behavior
Successfully process the dataset with map
### Environment info
datasets==3.3.1
|
open
|
https://github.com/huggingface/datasets/issues/7418
| 2025-02-21T10:58:06 | 2025-07-11T13:06:10 | null |
{
"login": "alexxchen",
"id": 15705569,
"type": "User"
}
|
[] | false |
[] |
2,866,868,922 | 7,417 |
set dev version
| null |
closed
|
https://github.com/huggingface/datasets/pull/7417
| 2025-02-20T17:45:29 | 2025-02-20T17:47:50 | 2025-02-20T17:45:36 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,866,862,143 | 7,416 |
Release: 3.3.2
| null |
closed
|
https://github.com/huggingface/datasets/pull/7416
| 2025-02-20T17:42:11 | 2025-02-20T17:44:35 | 2025-02-20T17:43:28 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,865,774,546 | 7,415 |
Shard Dataset at specific indices
|
I have a dataset of sequences, where each example in the sequence is a separate row in the dataset (similar to LeRobotDataset). When running `Dataset.save_to_disk` how can I provide indices where it's possible to shard the dataset such that no episode spans more than 1 shard. Consequently, when I run `Dataset.load_from_disk`, how can I load just a subset of the shards to save memory and time on different ranks?
I guess an alternative to this would be, given a loaded `Dataset`, how can I run `Dataset.shard` such that sharding doesn't split any episode across shards?
|
open
|
https://github.com/huggingface/datasets/issues/7415
| 2025-02-20T10:43:10 | 2025-02-24T11:06:45 | null |
{
"login": "nikonikolov",
"id": 11044035,
"type": "User"
}
|
[] | false |
[] |
2,863,798,756 | 7,414 |
Gracefully cancel async tasks
| null |
closed
|
https://github.com/huggingface/datasets/pull/7414
| 2025-02-19T16:10:58 | 2025-02-20T14:12:26 | 2025-02-20T14:12:23 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,860,947,582 | 7,413 |
Documentation on multiple media files of the same type with WebDataset
|
The [current documentation](https://huggingface.co/docs/datasets/en/video_dataset) on a creating a video dataset includes only examples with one media file and one json. It would be useful to have examples where multiple files of the same type are included. For example, in a sign language dataset, you may have a base video and a video annotation of the extracted pose. According to the WebDataset documentation, this should be able to be done with period separated filenames. For example:
```e39871fd9fd74f55.base.mp4
e39871fd9fd74f55.pose.mp4
e39871fd9fd74f55.json
f18b91585c4d3f3e.base.mp4
f18b91585c4d3f3e.pose.mp4
f18b91585c4d3f3e.json
...
```
If you can confirm that this method of including multiple media files works with huggingface datasets and include an example in the documentation, I'd appreciate it.
|
open
|
https://github.com/huggingface/datasets/issues/7413
| 2025-02-18T16:13:20 | 2025-02-20T14:17:54 | null |
{
"login": "DCNemesis",
"id": 3616964,
"type": "User"
}
|
[] | false |
[] |
2,859,433,710 | 7,412 |
Index Error Invalid Ket is out of bounds for size 0 for code-search-net/code_search_net dataset
|
### Describe the bug
I am trying to do model pruning on sentence-transformers/all-mini-L6-v2 for the code-search-net/code_search_net dataset using INCTrainer class
However I am getting below error
```
raise IndexError(f"Invalid Key: {key is our of bounds for size {size}")
IndexError: Invalid key: 1840208 is out of bounds for size 0
```
### Steps to reproduce the bug
Model pruning on the above dataset using the below guide
https://huggingface.co/docs/optimum/en/intel/neural_compressor/optimization#pruning
### Expected behavior
The modsl should be successfully pruned
### Environment info
Torch version: 2.4.1
Python version: 3.8.10
|
open
|
https://github.com/huggingface/datasets/issues/7412
| 2025-02-18T05:58:33 | 2025-02-18T06:42:07 | null |
{
"login": "harshakhmk",
"id": 56113657,
"type": "User"
}
|
[] | false |
[] |
2,858,993,390 | 7,411 |
Attempt to fix multiprocessing hang by closing and joining the pool before termination
|
https://github.com/huggingface/datasets/issues/6393 has plagued me on and off for a very long time. I have had various workarounds (one time combining two filter calls into one filter call removed the issue, another time making rank 0 go first resolved a cache race condition, one time i think upgrading the version of something resolved it). I don't know hf datasets well enough to fully understand the root cause, but I _think_ this PR fixes it.
Evidence: I have an LLM Foundry training yaml/script (datasets version 3.2.0) that results in a hang ~1/10 times (for a baseline for this testing, it was 2/36 runs that hung). I also reran with the latest datasets version (3.3.1) and got 4/36 hung. Installing datasets from this PR, I was able to successful run the script 144 times without a hang occurring. Assuming the base probability is 1/10, this should be more than enough times to have confidence it works.
After adding some logging, I could see that the code hung during the __exit__ of the mp pool context manager, after all shards had been processed, and the tqdm context manager had exited.
My best explanation: When multiprocessing pool __exit__ is called, it calls pool.terminate, which forcefully exits all the processes (and calls code related to this that I haven't looked at closely). I'm guessing this forceful termination has a bad interaction with some multithreading/multiprocessing that hf datasets does. If we instead call pool.close and pool.join before the pool.terminate happens, perhaps whatever that bad interaction is is able to complete gracefully, and then terminate call proceeds without issue.
If this PR seems good to you, I'd be very appreciative if you were able to do a patch release including it. Thank you!
@lhoestq
|
closed
|
https://github.com/huggingface/datasets/pull/7411
| 2025-02-17T23:58:03 | 2025-02-19T21:11:24 | 2025-02-19T13:40:32 |
{
"login": "dakinggg",
"id": 43149077,
"type": "User"
}
|
[] | true |
[] |
2,858,085,707 | 7,410 |
Set dev version
| null |
closed
|
https://github.com/huggingface/datasets/pull/7410
| 2025-02-17T14:54:39 | 2025-02-17T14:56:58 | 2025-02-17T14:54:56 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,858,079,508 | 7,409 |
Release: 3.3.1
| null |
closed
|
https://github.com/huggingface/datasets/pull/7409
| 2025-02-17T14:52:12 | 2025-02-17T14:54:32 | 2025-02-17T14:53:13 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,858,012,313 | 7,408 |
Fix filter speed regression
|
close https://github.com/huggingface/datasets/issues/7404
|
closed
|
https://github.com/huggingface/datasets/pull/7408
| 2025-02-17T14:25:32 | 2025-02-17T14:28:48 | 2025-02-17T14:28:46 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,856,517,442 | 7,407 |
Update use_with_pandas.mdx: to_pandas() correction in last section
|
last section ``to_pandas()"
|
closed
|
https://github.com/huggingface/datasets/pull/7407
| 2025-02-17T01:53:31 | 2025-02-20T17:28:04 | 2025-02-20T17:28:04 |
{
"login": "ibarrien",
"id": 7552335,
"type": "User"
}
|
[] | true |
[] |
2,856,441,206 | 7,406 |
Adding Core Maintainer List to CONTRIBUTING.md
|
### Feature request
I propose adding a core maintainer list to the `CONTRIBUTING.md` file.
### Motivation
The Transformers and Liger-Kernel projects maintain lists of core maintainers for each module.
However, the Datasets project doesn't have such a list.
### Your contribution
I have nothing to add here.
|
closed
|
https://github.com/huggingface/datasets/issues/7406
| 2025-02-17T00:32:40 | 2025-03-24T10:57:54 | 2025-03-24T10:57:54 |
{
"login": "jp1924",
"id": 93233241,
"type": "User"
}
|
[
{
"name": "enhancement",
"color": "a2eeef"
}
] | false |
[] |
2,856,372,814 | 7,405 |
Lazy loading of environment variables
|
### Describe the bug
Loading a `.env` file after an `import datasets` call does not correctly use the environment variables.
This is due the fact that environment variables are read at import time:
https://github.com/huggingface/datasets/blob/de062f0552a810c52077543c1169c38c1f0c53fc/src/datasets/config.py#L155C1-L155C80
### Steps to reproduce the bug
```bash
# make tmp dir
mkdir -p /tmp/debug-env
# make .env file
echo HF_HOME=/tmp/debug-env/data > /tmp/debug-env/.env
# first load dotenv, downloads to /tmp/debug-env/data
uv run --with datasets,python-dotenv python3 -c \
'import dotenv; dotenv.load_dotenv("/tmp/debug-env/.env"); import datasets; datasets.load_dataset("Anthropic/hh-rlhf")'
# first import datasets, downloads to `~/.cache/huggingface`
uv run --with datasets,python-dotenv python3 -c \
'import datasets; import dotenv; dotenv.load_dotenv("/tmp/debug-env/.env"); datasets.load_dataset("Anthropic/hh-rlhf")'
```
### Expected behavior
I expect that setting environment variables with something like this:
```python3
if __name__ == "__main__":
load_dotenv()
main()
```
works correctly.
### Environment info
"datasets>=3.3.0",
|
open
|
https://github.com/huggingface/datasets/issues/7405
| 2025-02-16T22:31:41 | 2025-02-17T15:17:18 | null |
{
"login": "nikvaessen",
"id": 7225987,
"type": "User"
}
|
[] | false |
[] |
2,856,366,207 | 7,404 |
Performance regression in `dataset.filter`
|
### Describe the bug
We're filtering dataset of ~1M (small-ish) records. At some point in the code we do `dataset.filter`, before (including 3.2.0) it was taking couple of seconds, and now it takes 4 hours.
We use 16 threads/workers, and stack trace at them look as follows:
```
Traceback (most recent call last):
File "/python/lib/python3.12/site-packages/multiprocess/process.py", line 314, in _bootstrap
self.run()
File "/python/lib/python3.12/site-packages/multiprocess/process.py", line 108, in run
self._target(*self._args, **self._kwargs)
File "/python/lib/python3.12/site-packages/multiprocess/pool.py", line 125, in worker
result = (True, func(*args, **kwds))
^^^^^^^^^^^^^^^^^^^
File "/python/lib/python3.12/site-packages/datasets/utils/py_utils.py", line 678, in _write_generator_to_queue
for i, result in enumerate(func(**kwargs)):
File "/python/lib/python3.12/site-packages/datasets/arrow_dataset.py", line 3511, in _map_single
for i, batch in iter_outputs(shard_iterable):
File "/python/lib/python3.12/site-packages/datasets/arrow_dataset.py", line 3461, in iter_outputs
yield i, apply_function(example, i, offset=offset)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/python/lib/python3.12/site-packages/datasets/arrow_dataset.py", line 3390, in apply_function
processed_inputs = function(*fn_args, *additional_args, **fn_kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/python/lib/python3.12/site-packages/datasets/arrow_dataset.py", line 6416, in get_indices_from_mask_function
indices_array = indices_mapping.column(0).take(indices_array)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 1079, in pyarrow.lib.ChunkedArray.take
File "/python/lib/python3.12/site-packages/pyarrow/compute.py", line 458, in take
def take(data, indices, *, boundscheck=True, memory_pool=None):
```
### Steps to reproduce the bug
1. Save dataset of 1M records in arrow
2. Filter it with 16 threads
3. Watch it take too long
### Expected behavior
Filtering done fast
### Environment info
datasets 3.3.0, python 3.12
|
closed
|
https://github.com/huggingface/datasets/issues/7404
| 2025-02-16T22:19:14 | 2025-02-17T17:46:06 | 2025-02-17T14:28:48 |
{
"login": "ttim",
"id": 82200,
"type": "User"
}
|
[] | false |
[] |
2,855,880,858 | 7,402 |
Fix a typo in arrow_dataset.py
|
"in the feature" should be "in the future"
|
closed
|
https://github.com/huggingface/datasets/pull/7402
| 2025-02-16T04:52:02 | 2025-02-20T17:29:28 | 2025-02-20T17:29:28 |
{
"login": "jingedawang",
"id": 7996256,
"type": "User"
}
|
[] | true |
[] |
2,853,260,869 | 7,401 |
set dev version
| null |
closed
|
https://github.com/huggingface/datasets/pull/7401
| 2025-02-14T10:17:03 | 2025-02-14T10:19:20 | 2025-02-14T10:17:13 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,853,098,442 | 7,399 |
Synchronize parameters for various datasets
|
### Describe the bug
[IterableDatasetDict](https://huggingface.co/docs/datasets/v3.2.0/en/package_reference/main_classes#datasets.IterableDatasetDict.map) map function is missing the `desc` parameter. You can see the equivalent map function for [Dataset here](https://huggingface.co/docs/datasets/v3.2.0/en/package_reference/main_classes#datasets.Dataset.map).
There might be other parameters missing - I haven't checked.
### Steps to reproduce the bug
from datasets import Dataset, IterableDataset, IterableDatasetDict
ds = IterableDatasetDict({"train": Dataset.from_dict({"a": range(6)}).to_iterable_dataset(num_shards=3),
"validate": Dataset.from_dict({"a": range(6)}).to_iterable_dataset(num_shards=3)})
for d in ds["train"]:
print(d)
ds = ds.map(lambda x: {k: v+1 for k, v in x.items()}, desc="increment")
for d in ds["train"]:
print(d)
### Expected behavior
The description parameter should be available for all datasets (or none).
### Environment info
- `datasets` version: 3.2.0
- Platform: Linux-6.1.85+-x86_64-with-glibc2.35
- Python version: 3.11.11
- `huggingface_hub` version: 0.28.1
- PyArrow version: 17.0.0
- Pandas version: 2.2.2
- `fsspec` version: 2024.9.0
|
open
|
https://github.com/huggingface/datasets/issues/7399
| 2025-02-14T09:15:11 | 2025-02-19T11:50:29 | null |
{
"login": "grofte",
"id": 7976840,
"type": "User"
}
|
[] | false |
[] |
2,853,097,869 | 7,398 |
Release: 3.3.0
| null |
closed
|
https://github.com/huggingface/datasets/pull/7398
| 2025-02-14T09:15:03 | 2025-02-14T09:57:39 | 2025-02-14T09:57:37 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,852,829,763 | 7,397 |
Kannada dataset(Conversations, Wikipedia etc)
| null |
closed
|
https://github.com/huggingface/datasets/pull/7397
| 2025-02-14T06:53:03 | 2025-02-20T17:28:54 | 2025-02-20T17:28:53 |
{
"login": "Likhith2612",
"id": 146451281,
"type": "User"
}
|
[] | true |
[] |
2,853,201,277 | 7,400 |
504 Gateway Timeout when uploading large dataset to Hugging Face Hub
|
### Description
I encountered consistent 504 Gateway Timeout errors while attempting to upload a large dataset (approximately 500GB) to the Hugging Face Hub. The upload fails during the process with a Gateway Timeout error.
I will continue trying to upload. While it might succeed in future attempts, I wanted to report this issue in the meantime.
### Reproduction
- I attempted the upload 3 times
- Each attempt resulted in the same 504 error during the upload process (not at the start, but in the middle of the upload)
- Using `dataset.push_to_hub()` method
### Environment Information
```
- huggingface_hub version: 0.28.0
- Platform: Linux-6.8.0-52-generic-x86_64-with-glibc2.39
- Python version: 3.11.10
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Running in Google Colab Enterprise ?: No
- Token path ?: /home/hotchpotch/.cache/huggingface/token
- Has saved token ?: True
- Who am I ?: hotchpotch
- Configured git credential helpers: store
- FastAI: N/A
- Tensorflow: N/A
- Torch: 2.5.1
- Jinja2: 3.1.5
- Graphviz: N/A
- keras: N/A
- Pydot: N/A
- Pillow: 10.4.0
- hf_transfer: N/A
- gradio: N/A
- tensorboard: N/A
- numpy: 1.26.4
- pydantic: 2.10.6
- aiohttp: 3.11.11
- ENDPOINT: https://huggingface.co
- HF_HUB_CACHE: /home/hotchpotch/.cache/huggingface/hub
- HF_ASSETS_CACHE: /home/hotchpotch/.cache/huggingface/assets
- HF_TOKEN_PATH: /home/hotchpotch/.cache/huggingface/token
- HF_STORED_TOKENS_PATH: /home/hotchpotch/.cache/huggingface/stored_tokens
- HF_HUB_OFFLINE: False
- HF_HUB_DISABLE_TELEMETRY: False
- HF_HUB_DISABLE_PROGRESS_BARS: None
- HF_HUB_DISABLE_SYMLINKS_WARNING: False
- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False
- HF_HUB_DISABLE_IMPLICIT_TOKEN: False
- HF_HUB_ENABLE_HF_TRANSFER: False
- HF_HUB_ETAG_TIMEOUT: 10
- HF_HUB_DOWNLOAD_TIMEOUT: 10
```
### Full Error Traceback
```python
Traceback (most recent call last):
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/huggingface_hub/utils/_http.py", line 406, in hf_raise_for_status
response.raise_for_status()
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/requests/models.py", line 1024, in raise_for_status
raise HTTPError(http_error_msg, response=self)
requests.exceptions.HTTPError: 504 Server Error: Gateway Time-out for url: https://huggingface.co/datasets/hotchpotch/fineweb-2-edu-japanese.git/info/lfs/objects/batch
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/create_edu_japanese_ds/upload_edu_japanese_ds.py", line 12, in <module>
ds.push_to_hub("hotchpotch/fineweb-2-edu-japanese", private=True)
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/datasets/dataset_dict.py", line 1665, in push_to_hub
split_additions, uploaded_size, dataset_nbytes = self[split]._push_parquet_shards_to_hub(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/datasets/arrow_dataset.py", line 5301, in _push_parquet_shards_to_hub
api.preupload_lfs_files(
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/huggingface_hub/hf_api.py", line 4215, in preupload_lfs_files
_upload_lfs_files(
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/huggingface_hub/_commit_api.py", line 395, in _upload_lfs_files
batch_actions_chunk, batch_errors_chunk = post_lfs_batch_info(
^^^^^^^^^^^^^^^^^^^^
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/huggingface_hub/lfs.py", line 168, in post_lfs_batch_info
hf_raise_for_status(resp)
File "/home/hotchpotch/src/github.com/hotchpotch/fineweb-2-edu-classifier-japanese/.venv/lib/python3.11/site-packages/huggingface_hub/utils/_http.py", line 477, in hf_raise_for_status
raise _format(HfHubHTTPError, str(e), response) from e
huggingface_hub.errors.HfHubHTTPError: 504 Server Error: Gateway Time-out for url: https://huggingface.co/datasets/hotchpotch/fineweb-2-edu-japanese.git/info/lfs/objects/batch
```
|
open
|
https://github.com/huggingface/datasets/issues/7400
| 2025-02-14T02:18:35 | 2025-02-14T23:48:36 | null |
{
"login": "hotchpotch",
"id": 3500,
"type": "User"
}
|
[] | false |
[] |
2,851,716,755 | 7,396 |
Update README.md
| null |
closed
|
https://github.com/huggingface/datasets/pull/7396
| 2025-02-13T17:44:36 | 2025-02-13T17:46:57 | 2025-02-13T17:44:51 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,851,575,160 | 7,395 |
Update docs
|
- update min python version
- replace canonical dataset names with new names
- avoid examples with trust_remote_code
|
closed
|
https://github.com/huggingface/datasets/pull/7395
| 2025-02-13T16:43:15 | 2025-02-13T17:20:32 | 2025-02-13T17:20:30 |
{
"login": "lhoestq",
"id": 42851186,
"type": "User"
}
|
[] | true |
[] |
2,847,172,115 | 7,394 |
Using load_dataset with data_files and split arguments yields an error
|
### Describe the bug
It seems the list of valid splits recorded by the package becomes incorrectly overwritten when using the `data_files` argument.
If I run
```python
from datasets import load_dataset
load_dataset("allenai/super", split="all_examples", data_files="tasks/expert.jsonl")
```
then I get the error
```
ValueError: Unknown split "all_examples". Should be one of ['train'].
```
However, if I run
```python
from datasets import load_dataset
load_dataset("allenai/super", split="train", name="Expert")
```
then I get
```
ValueError: Unknown split "train". Should be one of ['all_examples'].
```
### Steps to reproduce the bug
Run
```python
from datasets import load_dataset
load_dataset("allenai/super", split="all_examples", data_files="tasks/expert.jsonl")
```
### Expected behavior
No error.
### Environment info
Python = 3.12
datasets = 3.2.0
|
open
|
https://github.com/huggingface/datasets/issues/7394
| 2025-02-12T04:50:11 | 2025-02-12T04:50:11 | null |
{
"login": "devon-research",
"id": 61103399,
"type": "User"
}
|
[] | false |
[] |
2,846,446,674 | 7,393 |
Optimized sequence encoding for scalars
|
The change in https://github.com/huggingface/datasets/pull/3197 introduced redundant list-comprehensions when `obj` is a long sequence of scalars. This becomes a noticeable overhead when loading data from an `IterableDataset` in the function `_apply_feature_types_on_example` and can be eliminated by adding a check for scalars in `encode_nested_example` proposed here.
In the following code example
```
import time
from datasets.features import Sequence, Value
from datasets.features.features import encode_nested_example
schema = Sequence(Value("int32"))
obj = list(range(100000))
start = time.perf_counter()
result = encode_nested_example(schema, obj)
stop = time.perf_counter()
print(f"Time spent is {stop-start} sec")
```
`encode_nested_example` becomes 492x faster (from 0.0769 to 0.0002 sec), respectively 322x (from 0.00814 to 0.00003 sec) for a list of length 10000, on a GH200 system, making it unnoticeable when loading data with tokenization.
Another change is made to avoid creating arrays from scalars and afterwards re-extracting them during casting to python (`obj == obj.__array__()[()]` in that case), which avoids a regression in the array write benchmarks.
|
closed
|
https://github.com/huggingface/datasets/pull/7393
| 2025-02-11T20:30:44 | 2025-02-13T17:11:33 | 2025-02-13T17:11:32 |
{
"login": "lukasgd",
"id": 38319063,
"type": "User"
}
|
[] | true |
[] |
2,846,095,043 | 7,392 |
push_to_hub payload too large error when using large ClassLabel feature
|
### Describe the bug
When using `datasets.DatasetDict.push_to_hub` an `HfHubHTTPError: 413 Client Error: Payload Too Large for url` is raised if the dataset contains a large `ClassLabel` feature. Even if the total size of the dataset is small.
### Steps to reproduce the bug
``` python
import random
import sys
import datasets
random.seed(42)
def random_str(sz):
return "".join(chr(random.randint(ord("a"), ord("z"))) for _ in range(sz))
data = datasets.DatasetDict(
{
str(i): datasets.Dataset.from_dict(
{
"label": [list(range(3)) for _ in range(10)],
"abstract": [random_str(10_000) for _ in range(10)],
},
)
for i in range(3)
}
)
features = data["1"].features.copy()
features["label"] = datasets.Sequence(
datasets.ClassLabel(names=[str(i) for i in range(50_000)])
)
data = data.map(lambda examples: {}, features=features)
feat_size = sys.getsizeof(data["1"].features["label"].feature.names)
print(f"Size of ClassLabel names: {feat_size}")
# Size of ClassLabel names: 444376
data.push_to_hub("dconnell/pubtator3_test")
```
Note that this succeeds if `ClassLabel` has fewer names or if `ClassLabel` is replaced with `Value("int64")`
### Expected behavior
Should push the dataset to hub.
### Environment info
Copy-and-paste the text below in your GitHub issue.
- `datasets` version: 3.2.0
- Platform: Linux-5.15.0-126-generic-x86_64-with-glibc2.35
- Python version: 3.12.8
- `huggingface_hub` version: 0.28.1
- PyArrow version: 19.0.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.9.0
|
open
|
https://github.com/huggingface/datasets/issues/7392
| 2025-02-11T17:51:34 | 2025-02-11T18:01:31 | null |
{
"login": "DavidRConnell",
"id": 35470740,
"type": "User"
}
|
[] | false |
[] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.