id int64 600M 4.6B | node_id stringlengths 18 24 | number int64 2 8.24k | title stringlengths 1 290 | state stringclasses 2
values | state_reason stringclasses 5
values | created_at timestamp[s]date 2020-04-14 18:18:51 2026-06-05 18:08:17 | updated_at timestamp[s]date 2020-04-29 09:23:05 2026-06-05 21:18:34 | closed_at stringlengths 0 20 | html_url stringlengths 48 51 | user stringlengths 3 26 | labels listlengths 0 4 | assignees listlengths 0 4 | milestone stringclasses 7
values | comments listlengths 0 70 | author_association stringclasses 4
values | body stringlengths 0 228k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
4,599,059,760 | I_kwDODunzps8AAAABEiAVMA | 8,242 | "eval_strategy": "no" perform evaluation | open | 2026-06-05T18:08:17 | 2026-06-05T21:18:34 | https://github.com/huggingface/datasets/issues/8242 | SamuelLarkin | [] | [] | [] | NONE | ### Describe the bug
Hi,
I'm trying to train a diffusion model and I getting stuck with
```
strace -p 967873
strace: Process 967873 attached
ioctl(9, _IOC(_IOC_READ|_IOC_WRITE, 0x46, 0x2a, 0x20), 0x7fff18579d90) = 0
ioctl(9, _IOC(_IOC_READ|_IOC_WRITE, 0x46, 0x2a, 0x20), 0x7fff18579d90) = 0
```
But my issue is that ... | |||
4,597,756,669 | I_kwDODunzps8AAAABEgwy_Q | 8,241 | load_dataset("json", ...) infers a different schema when the JSONL file starts with a UTF-8 BOM | open | 2026-06-05T14:39:46 | 2026-06-05T21:03:25 | https://github.com/huggingface/datasets/issues/8241 | f-salvetti | [] | [] | [
"#self-assign"
] | NONE | ### Describe the bug
When a JSONL file starts with a UTF-8 BOM (`\xef\xbb\xbf`), `datasets.load_dataset("json", data_files=...)` infers a different schema than when the same file is loaded without the BOM. Specifically, the loader's "mixed-struct-types" pre-scan — which normally promotes columns with heterogeneous nes... | |||
4,501,694,591 | I_kwDODunzps8AAAABDFJofw | 8,217 | Multi-GPU Training with Large Dataset — NCCL Timeouts / Slow Data Loading + Excessive Memory Usage with keep_in_memory=True | closed | completed | 2026-05-22T09:55:13 | 2026-06-03T02:12:43 | 2026-06-03T02:12:43Z | https://github.com/huggingface/datasets/issues/8217 | D222097 | [] | [] | [
"Hi @D222097,\n\nWe have audited the architectural footprint of your multi-GPU training setup. The behaviors you are experiencing are not bugs within the `datasets` library, but rather fundamental limitations of Python's process architecture interacting with hardware topologies. \n\nHere is the clinical breakdown o... | NONE | Hi, I'm training a large dataset using Lerobot (which relies on Hugging Face Datasets) and encountering severe performance issues when scaling to multi-GPU training.
### Problem Description
- Unstable/slow data loading + NCCL timeouts
- Step time fluctuates drastically: ~1s/step (fast) vs ~45s/step (extremely slow)
... | |
4,460,017,531 | I_kwDODunzps8AAAABCdZ3ew | 8,204 | Please enable Private Vulnerability Reporting | open | 2026-05-16T13:28:01 | 2026-05-16T13:28:01 | https://github.com/huggingface/datasets/issues/8204 | 0xBassia | [] | [] | [] | NONE | Hi datasets team,
I have a security finding in datasets that I would like to disclose privately. GitHub Private Vulnerability Reporting is currently disabled on this repo.
Could you turn it on? It's at Settings > Security > Code security and analysis > Private vulnerability reporting. Once enabled I will file the ful... | |||
4,456,407,352 | I_kwDODunzps8AAAABCZ9hOA | 8,201 | Encountered an error when loading data in WebDataset format using `load_datasets` during multi-matchines training. | open | 2026-05-15T19:21:24 | 2026-05-20T12:58:57 | https://github.com/huggingface/datasets/issues/8201 | aihao2000 | [] | [] | [
"Hi @aihao2000 ! I was looking into this multi-machine training scenario and just opened a PR (#8203 ) that safely skips feature inference for empty shards during distributed streaming.\n\nplease! Check it out and see if it looks like it will resolve the weird bugs you were hitting!",
"it looks like there less da... | NONE | ### Describe the bug
Single-machine training works fine, but multi-machine training throws up all sorts of weird bugs. Help me!
### Steps to reproduce the bug
train_dataset = load_dataset(
"webdataset", data_files=args.train_dataset, split="train", streaming=True,cache_dir='/dev/shm/.cache'
)
args.train... | |||
4,455,786,605 | I_kwDODunzps8AAAABCZXobQ | 8,199 | Replace AutoFeatureExtractor with AutoImageProcessor in image preprocessing docs | open | 2026-05-15T17:34:53 | 2026-05-15T17:56:57 | https://github.com/huggingface/datasets/issues/8199 | ajaystar8 | [] | [] | [
"I'd like to work on this. I'll open a PR shortly."
] | NONE | ## Summary
The `docs/source/use_dataset.mdx` documentation uses `AutoFeatureExtractor` from `transformers` in its image preprocessing example. For vision models, `AutoImageProcessor` is now the recommended API and better reflects current `transformers` conventions.
## Problem
In the [[Apply data augmentations](https... | |||
4,452,221,355 | I_kwDODunzps8AAAABCV-Bqw | 8,197 | Feature Request / Discussion: Skip origin_metadata fetch when streaming=True? | open | 2026-05-15T07:43:57 | 2026-06-05T12:34:08 | https://github.com/huggingface/datasets/issues/8197 | yuxin00j | [] | [] | [
"I think it's great that it checks that every file does exist by default. It's relatively cheap since it uses fsspec which caches file info at folder level.\n\nOn the other hand why not add skip_origin_metadata=False from @TheDerbiedOne by default to DatasetBuilder and document what it does (load_dataset passes kwa... | NONE | Hi team!
I've been looking into the initialization process of the `DatasetBuilder` and noticed that the ETag/origin metadata caching path is executed unconditionally.
While fetching the `origin_metadata` and calculating the ETag is necessary for safely managing a local disk cache when using `as_dataset()`, it seems t... | |||
4,397,310,407 | I_kwDODunzps8AAAABBhmhxw | 8,178 | The full dataset viewer is not available (click to read why). Only showing a preview of the rows. Job manager crashed while running this job (missing heartbeats). Error code: JobManagerCrashedError | open | 2026-05-07T08:33:48 | 2026-05-31T00:09:38 | https://github.com/huggingface/datasets/issues/8178 | kakamond | [] | [] | [
"Oh yeah, we hit something pretty similar when running a dataset viewer on a large collection (~3M rows) a few months back. The JobManagerCrashedError with missing heartbeats usually showed up for us when the server process ran out of memory or couldn’t keep up with the load spikes during dataset streaming. Turns o... | NONE | ### Describe the bug
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Job manager crashed while running this job (missing heartbeats).
Error code: JobManagerCrashedError
see: https://huggingface.co/datasets/Genius-Society/tt100k/viewer
### Steps to reproduce the bug... | |||
4,376,879,740 | I_kwDODunzps8AAAABBOHifA | 8,171 | Dataset UI shows 0:00/0:00 for audio after push_to_hub | closed | completed | 2026-05-04T13:17:25 | 2026-05-20T10:46:58 | 2026-05-20T10:46:58Z | https://github.com/huggingface/datasets/issues/8171 | MNIKIEMA | [] | [] | [
"Hello,\nI am trying to find my first issue to resolve. I can't reproduce your problem. When I go in the hub UI to your dataset. I can see the length of each audio. It takes some time to load \n\n## Dataset Card\n<img width=\"584\" height=\"644\" alt=\"Image\" src=\"https://github.com/user-attachments/assets/8b0e38... | NONE | I’m seeing an issue where audio previews in the dataset viewer show `0:00/0:00` after uploading with push_to_hub.
## What I did
I tried to avoid `torchcodec` because I want to keep the packages small using the following code:
```python
import io
import os
import shutil
import tempfile
from pathlib import Path
imp... | |
4,366,520,018 | I_kwDODunzps8AAAABBEPO0g | 8,169 | Streaming dataset hangs consistently | closed | completed | 2026-05-01T20:27:58 | 2026-05-06T16:07:04 | 2026-05-06T16:06:08Z | https://github.com/huggingface/datasets/issues/8169 | michaelpginn | [] | [] | [
"+1 also encountered",
"I pushed a workaround for the current version of PyArrow (24.0.0) and older versions at https://github.com/huggingface/datasets/pull/8176\n\nFor future versions it should be fixed directly in PyArrow"
] | NONE | ### Describe the bug
When loading a dataset with `streaming=True`, some background process prevents the script from ever returning (or at least for a very long time).
This did not happen with `huggingface-hub < 1`.
### Steps to reproduce the bug
from datasets import load_dataset
ds = load_dataset("IRIIS-RESEARCH/N... | |
4,356,674,999 | I_kwDODunzps8AAAABA62Vtw | 8,167 | [BUG] Abnormal progress bar in dataset.map when load_from_cache_file=False | closed | completed | 2026-04-30T08:47:53 | 2026-05-29T10:59:07 | 2026-05-29T10:59:07Z | https://github.com/huggingface/datasets/issues/8167 | Jintao-Huang | [] | [] | [
"https://github.com/huggingface/datasets/blob/4e3bf52ff436d841e61211dfa096b5c47197ae36/src/datasets/arrow_dataset.py#L3593-L3598",
"Hi, can you provide more details on the issue ?",
"<img width=\"340\" height=\"26\" alt=\"Image\" src=\"https://github.com/user-attachments/assets/0dca6f47-4a8d-4871-a908-f58df856e... | NONE | ### Describe the bug
datasets 4.8.5
```python
from datasets import load_dataset, load_from_disk
data = load_from_disk('pretrain_cached_dataset/train')
def dummy_map(example):
return example
dataset = data.map(dummy_map)
dataset = data.map(dummy_map, load_from_cache_file=True) ... | |
4,354,878,889 | I_kwDODunzps8AAAABA5ItqQ | 8,165 | IterableDataset: corrupted EXIF image silently terminates streaming iterator instead of skipping the sample | open | 2026-04-30T02:20:54 | 2026-05-06T07:31:42 | https://github.com/huggingface/datasets/issues/8165 | LIUYellowBlack | [] | [] | [
"Hi ! Would it be acceptable to return None instead of skipping ? This way we can provide the same API for non-streaming (since the length of the dataset is known in advance and fixed). Then a subsequent .filter call can take care of skipping the samples. Something like this maybe ?\n\n```python\nds = load_dataset(... | NONE | ## Summary
When iterating over a streaming `IterableDataset` that contains images with corrupted EXIF metadata (e.g., a TIFF rational tag with denominator=0), PIL raises `ZeroDivisionError` inside `exif_transpose`. This exception propagates through the HuggingFace `datasets` pipeline and **terminates the streaming ite... | |||
4,354,825,277 | I_kwDODunzps8AAAABA5FcPQ | 8,164 | Lance: HF auth token dropped for private repos in streaming mode (broken `storage_options` lookup) | closed | completed | 2026-04-30T02:03:33 | 2026-04-30T15:38:43 | 2026-04-30T15:38:43Z | https://github.com/huggingface/datasets/issues/8164 | ericjaebeom | [] | [] | [
"I can see the fix. Changing `split(\"://\", 0)` to `split(\"://\", 1)` and removing the `+ \"://\"` suffix. The key lookup should be `\"hf\"` not `\"hf://\"`. \nI'd like to submit a PR for this. Is this still open?",
"Well, I am already writing a PR for this, and it is just about to be submitted."
] | CONTRIBUTOR | ### Describe the bug
When `Lance._split_generators` receives `hf://` URIs - i.e. anywhere a `StreamingDownloadManager` is in play (`load_dataset(..., streaming=True)`, `get_dataset_split_names`, and the dataset-viewer worker) - the HF auth token is silently dropped before reaching `lance.dataset(...)`. Lance then atte... | |
4,345,858,995 | I_kwDODunzps8AAAABAwiLsw | 8,159 | tests/conftest.py, tests/_test_patching.py, tests/fixtures/fsspec.py and other files are missing from the PYPI distribution | open | 2026-04-28T20:01:05 | 2026-04-30T14:40:23 | https://github.com/huggingface/datasets/issues/8159 | yurivict | [] | [] | [
"Hey, ran into something similar a while back when we were packaging a custom dataset loader for a client at v4.7.x. Turns out, PyPI distributions sometimes skip test files and fixtures like `conftest.py` since they're not critical for runtime—honestly, it's a packaging config thing in `setup.py` or `pyproject.toml... | NONE | ### Describe the bug
Version: 4.8.5
### Steps to reproduce the bug
n/a
### Expected behavior
n/a
### Environment info
n/a | |||
4,316,149,021 | I_kwDODunzps8AAAABAUM1HQ | 8,149 | [Optimization] Prevent per-thread instantiation of Cloud Storage FileSystem during Data loading initialization | open | 2026-04-23T13:09:19 | 2026-04-24T14:14:59 | https://github.com/huggingface/datasets/issues/8149 | ankitaluthra1 | [
"enhancement"
] | [] | [
"Is it open for contribution?",
"I think arbitrary filesystems would be recreated in every thread in any case, unlike HfFileSystem which has custom code to handle this."
] | NONE | ### Feature request
Modify the dataset loading initialization, so that fsspec filesystem instances (like GCSFileSystem or S3FileSystem) are instantiated once in the main thread and explicitly passed down to the background threads.
By pre-instantiating the fs object in the main thread (where the directory cache from g... | |||
4,306,478,911 | I_kwDODunzps8AAAABAK-nPw | 8,145 | `.map()` on a streaming IterableDataset silently skips examples after load_state_dict | closed | completed | 2026-04-22T03:08:39 | 2026-04-27T14:03:28 | 2026-04-27T14:03:28Z | https://github.com/huggingface/datasets/issues/8145 | MonliH | [] | [] | [
"I think this line here is causing issues: \nhttps://github.com/huggingface/datasets/blob/2724a653e7582077be647d3fdfe8cd1fac27b85f/src/datasets/iterable_dataset.py#L2225\nIt ignored force_convert_to_python, so it consumed a full Arrow chunk.\nI think it can be fixed if you add `self.force_convert_to_python` to the ... | NONE | ### Describe the bug
After `ds.map(fn)` on a streaming IterableDataset, saving state_dict() and loading it into a fresh dataset does not resume at the correct position. It jumps forward by up to one Arrow read-chunk (~1000 rows).
### Steps to reproduce the bug
```python
from datasets import Dataset
ds1 = Dataset.fr... | |
4,251,712,550 | I_kwDODunzps79a_wm | 8,131 | datasets.to_list: Behaviour change in version 4.7.0 with the addition of JSON type | closed | completed | 2026-04-13T06:06:43 | 2026-04-27T15:32:30 | 2026-04-27T15:32:30Z | https://github.com/huggingface/datasets/issues/8131 | ItsTania | [] | [] | [
"I think we can go for the same approach as in https://github.com/huggingface/datasets/pull/8122 which decodes the JSON data (and supports JSON nested in other objects)\n\n(Or decide to decode all the features types, which I think could make sense, but this needs more careful handling so this should be discussed in... | CONTRIBUTOR | ### Describe the bug
There is a change in the approach to decoding JSON types, added in https://github.com/huggingface/datasets/commit/d560b58e1f1ec96591a95a1780ee118f5eea6c74. Since 4.7.0, to_list() returns raw JSON strings for columns stored as Json(), while direct access via __getitem__ returns dicts.
In versions... | |
4,241,704,875 | I_kwDODunzps7800er | 8,129 | `load_dataset()` hangs when `hf_xet` is enabled | closed | completed | 2026-04-10T21:00:24 | 2026-04-10T21:07:56 | 2026-04-10T21:07:56Z | https://github.com/huggingface/datasets/issues/8129 | lewtun | [] | [] | [
"Apologies, seems to have been a stale cache issue. Nuking the datasets cache fixed the issue"
] | MEMBER | ### Describe the bug
Hi! I am getting a peculiar download issue on the latest versions of `datasets` and `hfh` where datasets hang indefinitely:
```py
from datasets import load_dataset
load_dataset("openai/gsm8k", "main", split="test")
```
Note that:
- streaming=True works.
- Direct Hub access works.
- Th... | |
4,223,104,906 | I_kwDODunzps77t3eK | 8,125 | Feature request: being able to disable chmod for flat permission filesystems | open | 2026-04-08T08:17:14 | 2026-04-08T08:22:42 | https://github.com/huggingface/datasets/issues/8125 | younesbelkada | [
"enhancement"
] | [] | [
"Context: the operation fails on GCS FUSE mounts, and likely will on similar s3-backed file systems.\n\n File \"/home/user/.venv/lib/python3.12/site-packages/datasets/arrow_dataset.py\", line 3747, in _map_single ... | NONE | ### Feature request
For flat permission file systems, `os.chmod` simply fails when trying to cache datasets - in some places such as: https://github.com/huggingface/datasets/blob/main/src/datasets/arrow_dataset.py#L4043-L4048 for arrow datasets.
This seems to be optional to obtained desired behavior as simply comment... | |||
4,201,197,701 | I_kwDODunzps76aTCF | 8,120 | Patch fields truncated — missing trailing newlines | closed | completed | 2026-04-03T16:07:47 | 2026-04-30T15:45:52 | 2026-04-30T15:45:52Z | https://github.com/huggingface/datasets/issues/8120 | hetao-git | [] | [] | [
"Hi, I'd like to work on this issue. I'll try to reproduce this issue and see where the newline is being stripped.",
"Hi @hetao-git\n\nI was able to reproduce this issue, I used the instance \"atlanhq_atlan-python_pr37\" from the AweAI-Team/Scale-SWE dataset. I observed that applying the GitHub patch (pull/37.dif... | NONE | Description: Multiple entries in the dataset have patches (test.patch and fix.patch) and f2p_patch fields with truncated content. The trailing blank lines that are part of valid unified diff format are stripped, causing git apply to fail with "corrupt patch" errors.
Affected instances (confirmed): atlanhq_atlan-python... | |
4,197,189,620 | I_kwDODunzps76LAf0 | 8,116 | `to_json()` serializes `Json()` feature as escaped string instead of native JSON object | closed | completed | 2026-04-02T21:59:15 | 2026-04-10T13:12:51 | 2026-04-10T13:12:51Z | https://github.com/huggingface/datasets/issues/8116 | ain-soph | [] | [] | [
"Hi! I ran into this issue as well, I think the root cause was that under `_batch_json()` of `JsonDatasetWriter` it was directly exporting the encoded string without decoding it first. I will try to add a PR with a small fix. ",
"#self-assign"
] | CONTRIBUTOR | ### Describe the bug
When a dataset column uses the `Json()` feature type, calling `to_json()` writes the value as an escaped JSON string (e.g., `"{\"cat_a\":1,\"cat_b\":2}"`) instead of a native JSON object (e.g., `{"cat_a":1,"cat_b":2}`).
This causes two problems:
1. The output JSON file is not interoperable with ... | |
4,176,918,326 | I_kwDODunzps749rc2 | 8,110 | `set_format` resets features | closed | completed | 2026-03-31T07:59:32 | 2026-05-11T07:15:16 | 2026-05-11T07:15:16Z | https://github.com/huggingface/datasets/issues/8110 | plutonium-239 | [] | [] | [
"couldn't replicate with: \n\n``` bash\nPlatform: macos\npython: 3.10.9\ndatasets: 4.8.4\nhuggingface_hub: 0.36.2\npyarrow: 23.0.1\npandas: 2.3.3\nfsspec: 2026.2.0\n\n```\n\n\n``` python\nimport datasets\nfrom datasets import Array2D, Dataset\nimport numpy as np\n\n# Create a dataset with Array2D feature\ndata = {\... | NONE | ### Describe the bug
Calling `dataset.set_format('...')` resets the informative features such as `Array2D` etc. back into `List(List(Value))` and so on.
### Steps to reproduce the bug
```py
>>> dataset.features
{'lld': Array2D(shape=(None, 26), dtype='float32')}
>>> dataset.set_format('torch')
>>> dataset.features
... | |
4,120,101,624 | I_kwDODunzps71k8L4 | 8,085 | ImportError: cannot import name 'VideoReader' from 'torchvision.io' when using datasets with torch format | closed | completed | 2026-03-23T11:16:20 | 2026-03-23T14:23:25 | 2026-03-23T14:00:57Z | https://github.com/huggingface/datasets/issues/8085 | DamJanusz | [] | [] | [
"`datasets` 4.8.4 is out and includes a fix :)"
] | NONE | ### Describe the bug
For PyTorch 2.11 + torchvision 0.26.0 when using the datasets library and setting the format to "torch", an ImportError occurs because the datasets library attempts to import VideoReader from torchvision.io, but this object is not found in the installed torchvision version.
**Error message**:
> ... | |
4,099,246,270 | I_kwDODunzps70VYi- | 8,079 | StepExamplesIterable._iter_arrow drops samples for non-zero ranks for single-sharded | closed | completed | 2026-03-19T04:38:51 | 2026-03-19T17:29:40 | 2026-03-19T17:29:40Z | https://github.com/huggingface/datasets/issues/8079 | wwwjn | [] | [
"lhoestq"
] | [] | NONE | ### Describe the bug
As titled
### Steps to reproduce the bug
Running following Python script:
```
"""
Reproducer for HuggingFace datasets 4.x bug:
StepExamplesIterable._iter_arrow drops samples for non-zero ranks
when Arrow batches have size 1 (e.g. single-shard streaming WebDataset).
Bug: split_dataset_by_node w... | |
4,086,560,531 | I_kwDODunzps7zk_cT | 8,078 | Inconsistent loading of LFS-hosted files in Dangindev/viet-cultural-vqa dataset | closed | completed | 2026-03-17T06:44:01 | 2026-03-30T04:14:29 | 2026-03-30T04:14:29Z | https://github.com/huggingface/datasets/issues/8078 | ngayngo9x | [] | [] | [
"I also have been stuck in same problem ",
"This should be fixed, thanks for reporting ! let us know if you ever encounter other issues in the future",
"It's okay. \n\nGood job, thanks your support! "
] | NONE | ### Describe the bug
Describe the bug
Some files in the Dangindev/viet-cultural-vqa dataset fail to load via the Hugging Face datasets library.
xet-hosted files load fine
LFS-hosted files sometimes fail
Example:
Fails: https://huggingface.co/datasets/Dangindev/viet-cultural-vqa/blob/main/images/am_thuc/banh_chung_T... | |
4,085,618,908 | I_kwDODunzps7zhZjc | 8,075 | `.batch()` error on formatted datasets | closed | completed | 2026-03-17T01:08:13 | 2026-04-10T13:14:03 | 2026-04-10T13:14:03Z | https://github.com/huggingface/datasets/issues/8075 | mariosasko | [] | [] | [
"Hi ! Good catch :) Since table-formatted iterable datasets have an Arrow path (i.e. they have `.iter_arrow()`) I guess a simple fix would be to make `.batch()` make the dataset arrow-formatted and use a `batch_arrow_fn` in that case (instead of the current `_batch_fn` that expects dictionaries). It would also be b... | COLLABORATOR | The `.batch()` method currently assumes the input (batch) is always a dictionary, which causes errors when it isn't. This can happen with formatted datasets, since formats like `"pyarrow"`, `"pandas"` (only affects `IterableDataset`), and `"polars"` return tables/dataframes instead of dictionaries.
For example:
```pyt... | |
4,073,606,584 | I_kwDODunzps7yzk24 | 8,065 | ARM audio support broken | open | 2026-03-13T22:06:36 | 2026-03-23T14:16:47 | https://github.com/huggingface/datasets/issues/8065 | leondz | [] | [] | [
"See also https://github.com/NVIDIA/garak/issues/1622#issuecomment-3937720381",
"Hi @leondz , thanks for the detailed bug report — the description and reproduction steps are really clear.\n\nI’d love to take a shot at investigating and potentially fixing this issue. Before I start digging into the codebase, I jus... | CONTRIBUTOR | ### Describe the bug
`datasets` 4.0 moved to using `torchcodec` for audio datasets, which doesn't support ARM hardware under Linux and Windows
### Steps to reproduce the bug
try:
* Try to install `torchcodec` under Linux ARM64
or:
* Try to install `torchcodec` under Windows ARM64
or, on either of those platforms... | |||
4,041,972,911 | I_kwDODunzps7w65yv | 8,052 | [Security] Zip/Rar/7z Archive Extraction Path Traversal (Incomplete CVE-2007-4559 Mitigation) | closed | completed | 2026-03-08T19:09:32 | 2026-03-10T23:27:45 | 2026-03-10T23:27:45Z | https://github.com/huggingface/datasets/issues/8052 | spartan8806 | [] | [] | [
"Hi! I'm currently exploring the codebase and came across this issue. I'd be interested in investigating it further and trying to reproduce the behavior locally.\n\nIf no one is actively working on it, I'd be happy to take a look and attempt a fix.\n",
"The steps to reproduce don't cause any file extraction outsi... | NONE | ## Security Vulnerability Report
**Reporter:** Conner Webber (conner.webber000@gmail.com)
**Severity:** HIGH (CVSS 3.1: 8.2)
**CWE:** CWE-22 (Improper Limitation of a Pathname to a Restricted Directory)
## Summary
The `ZipExtractor`, `RarExtractor`, and `SevenZipExtractor` classes in `datasets/utils/extract.py` call... | |
4,031,613,102 | I_kwDODunzps7wTYiu | 8,048 | Support 3D Datasets | closed | completed | 2026-03-06T01:23:08 | 2026-06-02T17:40:12 | 2026-06-02T17:38:55Z | https://github.com/huggingface/datasets/issues/8048 | Vinay-Umrethe | [
"enhancement"
] | [] | [
"Closing, as this got implemented in:\n\n[feat: add 3D mesh support and MeshFolder builder #8055](https://github.com/huggingface/datasets/pull/8055)\n\n[fix: embed_external_files=True for mesh support #8224](https://github.com/huggingface/datasets/pull/8224)"
] | CONTRIBUTOR | ### Feature request
In `datasets` library we have options to load `image` and `audio` datasets which can be viewed in the Dataset Viewer.
LIKE:
```bash
from datasets import load_dataset, Image
from datasets import load_dataset, Audio
```
> https://huggingface.co/docs/datasets/en/image_load & https://huggingface.co/do... | |
4,030,348,328 | I_kwDODunzps7wOjwo | 8,046 | `Dataset.map()` with `num_proc > 1` fails when some shards produce `List(Value('null'))` (empty lists) | closed | completed | 2026-03-05T20:12:13 | 2026-03-09T17:52:13 | 2026-03-09T17:52:13Z | https://github.com/huggingface/datasets/issues/8046 | ain-soph | [] | [] | [
"\n### Root cause\n\nIn `src/datasets/features/features.py`, both `_check_if_features_can_be_aligned` and `_align_features` use:\n\n```python\nisinstance(v, Value) and v.dtype == \"null\"\n```\n\nto check for null compatibility. However, `List(Value('null'))` is **not** an instance of `Value` — it's an instance of ... | CONTRIBUTOR | ### Describe the bug
When using `Dataset.map()` with `num_proc > 1`, if the map function produces a list column where some shards contain only empty lists `[]` and other shards contain non-empty lists of structs, the feature type inference is inconsistent across shards:
- Shards with only empty lists → `List(Value('n... | |
4,020,184,168 | I_kwDODunzps7vnyRo | 8,038 | Typo in iterable_dataset.py | closed | completed | 2026-03-04T05:37:15 | 2026-03-09T17:12:54 | 2026-03-09T17:12:54Z | https://github.com/huggingface/datasets/issues/8038 | sybaik1 | [] | [] | [] | NONE | ### Describe the bug
`streaming=True**kwargs,` should be `streaming=True, **kwargs,`
https://github.com/huggingface/datasets/blob/81027be09d5cd9f06a6d64ef1a8a3e9ebd0f86fd/src/datasets/iterable_dataset.py#L2950
### Steps to reproduce the bug
```
from Datasets import Dataset
IterableDataset.from_csv("file.csv")
```
#... | |
4,008,928,790 | I_kwDODunzps7u82YW | 8,037 | IterableDataset.filter() chaining fails due to features being lost when is_typed=True | closed | completed | 2026-03-02T02:30:24 | 2026-03-09T17:44:34 | 2026-03-09T17:44:34Z | https://github.com/huggingface/datasets/issues/8037 | sh0416 | [] | [] | [] | NONE | ### Describe the bug
Chaining multiple .filter() calls on an IterableDataset raises a TypeError: 'NoneType' object is not a mapping on the second .filter() call.
In `IterableDataset.filter()`, the internal `ex_iterable` is wrapped with FormattedExamplesIterable:
```python
ex_iterable = FormattedExamplesIterable(
... | |
4,004,188,523 | I_kwDODunzps7uqxFr | 8,034 | [BUG] load_datasets() cannot use the generated Arrow cache correctly | open | 2026-02-28T08:11:23 | 2026-03-04T13:50:08 | https://github.com/huggingface/datasets/issues/8034 | Nexround | [] | [] | [
"PYTHONHASHSEED=1 uv run python -c\n \"\nfrom datasets.load import LocalDatasetModuleFactory\nmod = LocalDatasetModuleFactory('/cache/datasets/imagenet-1k')\ndm = mod.get_module()\nprint('Hash:', dm.hash)\n\"\nResolving data files: 100%|█████████| 294/294 [00:00<00:00, 20584.68it/s]\nResolving data files: 100%|████... | CONTRIBUTOR | ### Describe the bug
The datasets library cannot use the generated Arrow cache correctly, seemingly due to a flaw in the internal hash symbol calculation logic.
The following code provides verification.
I am trying to locate the specific code position, and if there are further developments, I will update here.
### S... | |||
4,001,742,455 | I_kwDODunzps7uhb53 | 8,029 | `cache_dir` option in `download_config` in `load_dataset` is not respected | open | 2026-02-27T16:05:30 | 2026-02-27T23:34:28 | https://github.com/huggingface/datasets/issues/8029 | TsXor | [] | [] | [
"Actually, I used debugger to step into the code of `datasets`, and found `load_dataset` calls into `cached_path`, where it calls into `hf_hub_download` method for `hf://` links:\nhttps://github.com/huggingface/datasets/blob/6e8eaeed4d202b72eff44ba1a9ec5b7d81d2d3e6/src/datasets/utils/file_utils.py#L182-L195\n\nSome... | NONE | ### Describe the bug
Downloaded files still go to `~/.cache/huggingface/hub/` even if I specified `cache_dir` option in `download_config` in `load_dataset`.
### Steps to reproduce the bug
Run my freshly written script and found that downloaded files did not go where I want.
```python
'''
下载OpenWebText数据集，允许使用代理
'''
... | |||
3,974,281,900 | I_kwDODunzps7s4rqs | 8,015 | Iterating a streaming dataset correctly in random order | open | 2026-02-22T11:34:06 | 2026-05-11T15:08:20 | https://github.com/huggingface/datasets/issues/8015 | unrealwill | [] | [] | [
"Hi, reshuffling is actually pretty easy:\n\n```python\n\"\"\" shuffle.py \"\"\"\nfrom datasets import load_dataset\n\nnum_proc = 8 # tweak it here depending on your machine\nseed = 42\nds = load_dataset(\"jackyhate/text-to-image-2M\", num_proc=num_proc)\nds = ds.shuffle(seed=seed)\nds.push_to_hub(\"my-username/te... | NONE | ### Describe the bug
For training to work properly it is considered good practice in machine learning to sample the dataset randomly (uniformly). I am not sure how the user is expected to do it correctly when the creator of the dataset did a bad job.
https://huggingface.co/docs/hub/datasets-webdataset
```
**Shuffle*... | |||
3,963,855,875 | I_kwDODunzps7sQ6QD | 8,012 | HC3 dataset fails to load with datasets>=3 due to legacy script (HC3.py) | closed | completed | 2026-02-19T16:27:51 | 2026-02-20T11:58:20 | 2026-02-20T11:58:20Z | https://github.com/huggingface/datasets/issues/8012 | navan0 | [] | [] | [
"This is not a bug. This is intended behaviour. The only ways to resolve this are\n1. use the old `datasets` versions\n2. load the dataset and upload it to your own repo without a script\n3. raise an issue in the dataset and wait for the owner to update it."
] | NONE | ### Describe the bug
It looks like the current documentation suggests loading the dataset like this:
`load_dataset("Hello-SimpleAI/HC3", split="train")
`
But with datasets>=3.x, this raises:
RuntimeError: Dataset scripts are no longer supported, but found HC3.py
The dataset loads correctly if we specify the parqu... | |
3,963,448,558 | I_kwDODunzps7sPWzu | 8,010 | audiofolder / folder-based loader rejects metadata.csv when file_name column is inferred as large_string | closed | completed | 2026-02-19T15:06:09 | 2026-02-19T15:29:41 | 2026-02-19T15:29:41Z | https://github.com/huggingface/datasets/issues/8010 | Koosh0610 | [] | [] | [
"Proposing a PR for this [here](https://github.com/huggingface/datasets/pull/8011)\n\nPS: I didn't find a contributing documentation. Please recommend any changes in issue or PR if any. Thank you!"
] | NONE | ### Describe the bug
When loading a dataset with load_dataset("audiofolder", data_dir=...) and a metadata.csv (with a file_name column), loading can fail with:
`ValueError: file_name or *_file_name must be present as dictionary key (with type string) in metadata files`
even when the CSV clearly has a file_name colum... | |
3,946,695,329 | I_kwDODunzps7rPcqh | 8,007 | Add option for loading audio with video | open | 2026-02-16T09:11:02 | 2026-05-21T15:37:30 | https://github.com/huggingface/datasets/issues/8007 | Samoed | [] | [] | [
"Hi,\nI’m interested in working on this issue.\n\nFrom what I understand, audio decoding would involve an AudioDecoder, and previous discussions have suggested handling it separately. I’d like to explore whether integrating audio extraction in relation to the current video decoding workflow would be appropriate, or... | NONE | ### Describe the bug
Currently, `torchcodec` don't allow extracting `Audio` from `Video` https://github.com/meta-pytorch/torchcodec/issues/1158, but when I upload videos with audio to hub using `videofolder`, then this is not possible to retrieve audio from it. Probably `VideoDecoder` can be extended with `audio` para... | |||
3,944,394,074 | I_kwDODunzps7rGq1a | 8,006 | Regression: from_generator crashes if one generator call returns no results | closed | completed | 2026-02-15T16:16:48 | 2026-02-24T13:58:12 | 2026-02-24T13:58:12Z | https://github.com/huggingface/datasets/issues/8006 | hartmans | [] | [] | [] | CONTRIBUTOR | ### Describe the bug
Dataset.from_generator splits any list kwarg to the generator and makes a separate call to the generator for each member of the list even in the num_proc=1 case.
I have a generator that processes a number of files, filtering them and producing examples. It used to be the case that things worked fi... | |
3,941,908,297 | I_kwDODunzps7q9L9J | 8,005 | Multi-channel audio is automatically cast to mono, num_channels is ignored | open | 2026-02-14T17:28:03 | 2026-02-25T15:17:18 | https://github.com/huggingface/datasets/issues/8005 | ZackHodari | [] | [] | [
"**Workaround**\nDirectly load audio using torchcodec, this is what datasets does under the hood (but doesn't maintain multi-channel)\n\n```python\nimport torchcodec\n\ndecoder = torchcodec.decoders.AudioDecoder(audio[\"bytes\"])\naudio_samples = decoder.get_all_samples()\n\naudio = audio_samples.data.numpy()\nsamp... | NONE | ### Describe the bug
The `num_channels` parameter in `datasets.Audio()` is documented to preserve stereo channels when set to `None` (preserve original) or `2` (explicit stereo), but it currently downmixes all audio to mono regardless of this setting.
### Steps to reproduce the bug
```python
import numpy as np
impor... | |||
3,937,013,814 | I_kwDODunzps7qqhA2 | 8,002 | The `Sequence` class in features do not have "dtype" | open | 2026-02-13T12:45:19 | 2026-02-13T12:52:36 | https://github.com/huggingface/datasets/issues/8002 | gonzalo-santamaria-iic | [] | [] | [] | NONE | ### Describe the bug
**I'm not sure if this is a bug.**
I see that a `FeatureType` object contains an attribute called `self.dtype` that is not covered when this feature is a `Sequence` or a `List`.
When I try to run a multilabel classification with this example script from the transformers library:
https://github.... | |||
3,915,367,642 | I_kwDODunzps7pX8Ta | 7,999 | Too many dataloader workers: 4 (max is dataset.num_shards=3). Stopping 1 dataloader workers. | open | 2026-02-09T09:26:37 | 2026-02-15T14:25:56 | https://github.com/huggingface/datasets/issues/7999 | D222097 | [] | [] | [
"Hi, thanks for the clear question and code snippet!\n\nFrom my understanding, **hf_dataset.num_shards** represents the number of actual iterable partitions that the streaming dataset can be read from in parallel. This is not the same as the number of underlying Parquet files—many files can be grouped into a much s... | NONE | Hi !
I’m working on training with a large-scale dataset (100+ Parquet files) using lazy loading, and I’m struggling to understand/optimize the num_shards setting— in the lerobot repo: streaming_datasets.py:
```
from datasets import load_dataset
self.hf_dataset: datasets.IterableDataset = load_dataset(
self.... | |||
3,912,624,238 | I_kwDODunzps7pNehu | 7,998 | [doc] Inconsistant ENV VAR Name for Progress Bar Toggle | closed | completed | 2026-02-08T12:16:44 | 2026-02-16T15:29:50 | 2026-02-16T15:29:50Z | https://github.com/huggingface/datasets/issues/7998 | Moenupa | [] | [] | [] | NONE | Code uses env var name `HF_DATASETS_DISABLE_PROGRESS_BARS`.
https://github.com/huggingface/datasets/blob/025593f2f0722f31fc136e0ae45da4ff44d4416a/src/datasets/config.py#L221-L226
Docstrings and warnings report env var name `HF_DATASETS_DISABLE_PROGRESS_BAR` without the ending `S`.
https://github.com/huggingface/data... | |
3,906,330,806 | I_kwDODunzps7o1eC2 | 7,994 | Bump fsspec upper bound constraint | closed | completed | 2026-02-06T11:37:54 | 2026-02-16T15:29:00 | 2026-02-16T15:29:00Z | https://github.com/huggingface/datasets/issues/7994 | hadim | [] | [] | [] | NONE | Would it be possible to bump fsspec upper bound to the latest (2026.2.0)?
I saw you had some API compat issues in the past (https://github.com/huggingface/datasets/issues/7326) and I understand the need for an upper bound.
But I wonder if you think your CI and tests are a good proxy to catch fsspec API breakage? If t... | |
3,896,884,513 | I_kwDODunzps7oRb0h | 7,991 | list(api.list_datasets()) giving jsondecode error | closed | completed | 2026-02-04T14:39:46 | 2026-02-05T10:30:09 | 2026-02-05T10:30:09Z | https://github.com/huggingface/datasets/issues/7991 | Moll-j | [] | [] | [] | NONE | i am using the python api wrapper to list all datasets available on hugging face. This is for research, and i need all datasets to determine which % have language tags and other related questions requiring the total list. However, the following code that worked a few months ago:
```
from huggingface_hub import HfApi
a... | |
3,895,870,826 | I_kwDODunzps7oNkVq | 7,990 | Dataset.map crashes when first examples return None and later examples return dict — writer not initialized | open | 2026-02-04T10:43:20 | 2026-05-26T19:09:11 | https://github.com/huggingface/datasets/issues/7990 | meta-program | [] | [] | [
"map() expects a function that always return dict or that always return none",
"Hi, I'd like to work on this if the team is open to it.\n\nI understand from @lhoestq's comment that `map()` currently expects\na consistent return type. However, the crash with `AttributeError: \n'NoneType' object has no attribute 'w... | NONE | ### Describe the bug
I detected a serious [bug from datasets/arrow_dataset.py](https://github.com/huggingface/datasets/blob/main/src/datasets/arrow_dataset.py#L3676)
---
**Description of the bug**
`Dataset.map` crashes with `writer is None` when the map function returns `None` for the first few examples and a diction... | |||
3,895,353,435 | I_kwDODunzps7oLmBb | 7,988 | `Dataset.map()` breaks when `function` calls `import polars as pl` and `num_proc`>1: "UnboundLocalError: cannot access local variable 'pl' where it is not associated with a value" | closed | completed | 2026-02-04T08:42:23 | 2026-02-24T14:08:16 | 2026-02-24T14:08:16Z | https://github.com/huggingface/datasets/issues/7988 | ligz08 | [] | [] | [] | NONE | ### Describe the bug
# Repro
These two conditions seem to consistently reproduce the issue:
- function passed to `Dataset.map()` explicitly or implicitly calls `import polars as pl`
- `num_proc` > 1
# Trace
```
RemoteTraceback Traceback (most recent call last)
RemoteTraceback:
"""
Traceback... | |
3,892,776,651 | I_kwDODunzps7oBw7L | 7,986 | `Dataset.map()` causes cache miss/fingerprint change when closure captures self containing non-deterministic state. | open | 2026-02-03T19:16:49 | 2026-02-06T08:38:13 | https://github.com/huggingface/datasets/issues/7986 | Cloud0310 | [] | [] | [
"I suggest metion this in docs specifically for attention with use, tell users explicitly to pass arguments with `fn_kwargs` param or using `functools.partial` to create a pure funcion."
] | NONE | ### Describe the bug
When using `.map()` with a function defined inside a **class (of which has any non-deterministic states)** method (a closure), if that function captures `self` to access a configuration variable (e.g., self.foo), the fingerprint mechanism serializes the entire class instance state.
If the class i... | |||
3,891,431,105 | I_kwDODunzps7n8obB | 7,984 | Data | open | 2026-02-03T14:01:48 | 2026-02-03T14:01:48 | https://github.com/huggingface/datasets/issues/7984 | iLenceJhay | [] | [] | [] | NONE | ||||
3,869,700,866 | I_kwDODunzps7mpvMC | 7,970 | cast_column(..., Audio) fails with load_dataset("csv",) | open | 2026-01-29T09:33:35 | 2026-04-04T12:45:45 | https://github.com/huggingface/datasets/issues/7970 | jstangroome | [] | [] | [
"The following code *does* work:\n```py\nfrom datasets import load_dataset,Audio,Features\n\ndataset = load_dataset(\"csv\",data_files=\"audio.csv\",features=Features({\"audio\": Audio()}))\nprint(dataset[\"train\"][0][\"audio\"])\n```",
"Thanks for reporing ! Are you using pandas v3 by any chance ? The CSV loade... | NONE | ### Describe the bug
Attempt to load a dataset from a csv with a single `audio` column with a single row with a path to an audio file fails when casting the column to Audio, but the exact same dataset created from a dictionary succeeds.
### Steps to reproduce the bug
1. Have any valid audio file `audio.wav`
2. Have ... | |||
3,864,988,355 | I_kwDODunzps7mXwrD | 7,968 | Potential conflicting type checks and dead code in `/src/datasets/table.py` | closed | completed | 2026-01-28T11:34:53 | 2026-05-11T11:04:04 | 2026-05-11T11:04:04Z | https://github.com/huggingface/datasets/issues/7968 | rc4typecheck | [] | [] | [
"ConcatenationTable is a subclass of datasets.table.Table but not pa.Table, so it should be fine",
"Hi! Triaging stale issues — I don't think this one is actually dead code; the premise is off.\n\nThe claim is that `isinstance(table, ConcatenationTable)` at line 4 is unreachable because the line-2 `isinstance(tab... | NONE | When statically analyzing and manually reviewing the code, I noticed a potential logic conflicting in `/src/datasets/table.py` as follows:
```
def to_blocks(table: Union[pa.Table, Table]) -> list[list[TableBlock]]:
if isinstance(table, pa.Table):
return [[InMemoryTable(table)]]
elif isinstance(ta... | |
3,858,483,549 | I_kwDODunzps7l-8ld | 7,965 | `huggingface_hub.errors.HfHubHTTPError: 404 Client Error: Not Found for url` when fetching a dataset with `datasets.load_dataset` | closed | completed | 2026-01-27T02:20:31 | 2026-01-28T15:14:50 | 2026-01-28T15:14:50Z | https://github.com/huggingface/datasets/issues/7965 | harupy | [] | [] | [
"Hi ! Yes you should use `cornell-movie-review-data/rotten_tomatoes` instead of `rotten_tomatoes`, which is the legacy name. Those datasets have been moved under their actual owners accounts some time ago (but we were keeping the old names as aliases)\n\nSome other impacted names are:\n- `imdb` -> `stanfordnlp/imdb... | NONE | Not a bug but a question. We started getting the following error:
https://github.com/mlflow/mlflow/actions/runs/21368603305/job/61506951617
```
ests/data/test_huggingface_dataset_and_source.py::test_from_huggingface_dataset_constructs_expected_dataset_with_revision - huggingface_hub.errors.HfHubHTTPError: 404 Client ... | |
3,847,184,392 | I_kwDODunzps7lT2AI | 7,958 | [CUDA Tensors Not working in ~v4.5.0] set_format(type="torch", device="cuda") returns cpu | closed | completed | 2026-01-23T12:06:48 | 2026-03-04T13:17:28 | 2026-03-04T13:17:28Z | https://github.com/huggingface/datasets/issues/7958 | ai-nikolai | [] | [] | [
"@lhoestq tagging you here as you were on the previous issue, hope that's fine. ",
"Thanks for reporting, let me take a look",
"I reverted this change which caused the issue #7961 , I'll do a new release soon but in the meantime feel free to install `datasets` from source",
"@lhoestq thanks a lot. I am actual... | NONE | ### Describe the bug
The problem is that when calling:
```ds.set_format(type="torch", columns = ["input", "labels"], device="cuda")```
The device type of the individual datapoints is now: `cpu` as opposed to `cuda:0`.
With `v4.0.0` it still works. With `v4.5.0` it doesn't work anymore.
Related Issue:
https://github... | |
3,839,082,498 | I_kwDODunzps7k08AC | 7,956 | Is the 10k files / folder limit a hard limit for a dataset repo? | open | 2026-01-21T16:37:38 | 2026-01-22T09:38:38 | https://github.com/huggingface/datasets/issues/7956 | pavanramkumar | [
"enhancement"
] | [] | [
"Yes, that's a hard limit. Can you split your files into different folders? Or we'll probably have a new repo type in the near to mid future that will relax this limit a bit. ",
"Thanks! Working around this with a different sharding parameter to have fewer overall fragments (and therefore fewer files in `*.lance/... | NONE | ### Feature request
Can the hard limit of 10k files per folder be extended with acceptable loss in performance?
### Motivation
I'm uploading a lance dataset to huggingface hub and have a folder inside lance internals (`data/*.lance/_transactions`) that has > 20k atomic transaction records and my commits are being re... | |||
3,837,020,089 | I_kwDODunzps7ktEe5 | 7,954 | all_exhausted_without_replacement working same as first_exhausted | closed | completed | 2026-01-21T07:50:31 | 2026-03-23T14:03:13 | 2026-03-23T14:03:13Z | https://github.com/huggingface/datasets/issues/7954 | prathamk-tw | [] | [] | [
"https://github.com/huggingface/datasets/pull/7955",
"it was fixed by https://github.com/huggingface/datasets/pull/7955"
] | CONTRIBUTOR | ### Describe the bug
>>> from datasets import Dataset, interleave_datasets
>>> d1 = Dataset.from_dict({"a": [0, 1, 2]})
>>> d2 = Dataset.from_dict({"a": [10, 11, 12, 13]})
>>> d3 = Dataset.from_dict({"a": [20, 21, 22]})
>>> dataset = interleave_datasets([d1, d2, d3], stopping_strategy="all_exhausted_without_replacemen... |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- 22