Error code:   ConfigNamesError
Exception:    ValueError
Message:      Split name should match '^\w+(\.\w+)*$'' but got '{'caltech101-train', 'renderedsst2-train', 'stanfordcars-train', 'birdsnap-train', 'patchcamelyon-train', 'sun397-train', 'scimmir-train', 'country211-train', 'clevr_count-train', 'imagenet1k-train', 'cifar100-train', 'fer2013-train', 'cifar10-train', 'mnist-train', 'oxfordpets-train', 'clevr-train', 'eurosat-train', 'dtd-train', 'ucf101-train', 'food101-train', 'stl10-train', 'resisc45-train', 'gtsrb-train'}'.
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
                  config_names = get_dataset_config_names(
                      path=dataset,
                      token=hf_token,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
                  dataset_module = dataset_module_factory(
                      path,
                  ...<4 lines>...
                      **download_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1215, in dataset_module_factory
                  raise e1 from None
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1190, in dataset_module_factory
                  ).get_module()
                    ~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 648, in get_module
                  patterns = get_data_patterns(base_path, download_config=self.download_config)
                File "/usr/local/lib/python3.14/site-packages/datasets/data_files.py", line 493, in get_data_patterns
                  return _get_data_files_patterns(resolver)
                File "/usr/local/lib/python3.14/site-packages/datasets/data_files.py", line 279, in _get_data_files_patterns
                  raise ValueError(f"Split name should match '{_split_re}'' but got '{splits}'.")
              ValueError: Split name should match '^\w+(\.\w+)*$'' but got '{'caltech101-train', 'renderedsst2-train', 'stanfordcars-train', 'birdsnap-train', 'patchcamelyon-train', 'sun397-train', 'scimmir-train', 'country211-train', 'clevr_count-train', 'imagenet1k-train', 'cifar100-train', 'fer2013-train', 'cifar10-train', 'mnist-train', 'oxfordpets-train', 'clevr-train', 'eurosat-train', 'dtd-train', 'ucf101-train', 'food101-train', 'stl10-train', 'resisc45-train', 'gtsrb-train'}'.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

mieb-train-wds

WebDataset build of the MIEB zero-shot image-classification datasets, re-captioned so the class labels work as CLIP training text. Parquet twin: PumeTu/mieb-train.

4,434,196 samples · 493 shards · 23 datasets

Each shard member set is <key>.jpg|png|webp + <key>.txt (the caption) + <key>.cls (source class id). All image extensions fall inside OpenCLIP's default image key jpg;png;jpeg;webp, so no --image-key override is needed. Source bytes are passed through unmodified — no re-encoding.

Training with OpenCLIP

python -m open_clip_train.main \
  --dataset-type webdataset \
  --train-data "$(python -c "
from mieb_datasets.openclip import hub_wds_flags
print(hub_wds_flags('PumeTu/mieb-train-wds')['train_data'])")" \
  --train-num-samples 4434196

--train-num-samples is required: get_dataset_size() locates sizes.json with os.path.exists, which never succeeds for an https:// shard. data/sizes.json is published anyway so a downloaded copy is self-describing.

Rebalancing

Shards are named per dataset, so each becomes one :: component and can be weighted individually. Unweighted, ucf101 + imagenet1k take 69% of sampling — and ucf101's 1.79M rows are video frames from only ~13k clips, so that share is far more redundant than the row count suggests.

from mieb_datasets.openclip import hub_wds_flags
f = hub_wds_flags("PumeTu/mieb-train-wds", balance="sqrt")
# --train-data f["train_data"] --train-data-weights f["train_data_weights"] --dataset-resampled

Weights apply only with --dataset-resampled. OpenCLIP resamples shards, so a dataset's realised share is proportional to rows x weight.

dataset	rows	shards	share (none)	share (sqrt)	source
`ucf101`	1,786,096	179	40.28%	19.79%	`mteb/ucf101` @ `e0618988`
`imagenet1k`	1,281,167	134	28.89%	16.76%	`mteb/wds_imagenet1k` @ `b54c9af9`
`scimmir`	498,279	55	11.24%	10.45%	`mteb/SciMMIR` @ `2a10b6b1`
`patchcamelyon`	262,144	27	5.91%	7.58%	`mteb/wds_vtab-pcam` @ `6544d35f`
`sun397`	76,127	13	1.72%	4.08%	`mteb/sun397` @ `c684cff3`
`food101`	75,750	8	1.71%	4.07%	`mteb/food101` @ `7cca05bf`
`clevr`	63,000	7	1.42%	3.72%	`mteb/wds_vtab-clevr_closest_object_distance` @ `d2777bb7`
`clevr_count`	63,000	7	1.42%	3.72%	`mteb/wds_vtab-clevr_count_all` @ `2e31935f`
`mnist`	60,000	6	1.35%	3.63%	`mteb/mnist` @ `cf6afcba`
`cifar10`	50,000	5	1.13%	3.31%	`mteb/cifar10` @ `69a62dd1`
`cifar100`	50,000	5	1.13%	3.31%	`mteb/cifar100` @ `ac5511f8`
`country211`	31,650	6	0.71%	2.63%	`mteb/wds_country211` @ `c3875324`
`fer2013`	28,709	3	0.65%	2.51%	`mteb/wds_fer2013` @ `d9cabab3`
`gtsrb`	26,640	3	0.60%	2.42%	`mteb/wds_gtsrb` @ `a8f4bb6c`
`resisc45`	18,900	2	0.43%	2.04%	`mteb/resisc45` @ `a5b7c0d4`
`eurosat`	16,200	2	0.37%	1.88%	`mteb/eurosat-rgb` @ `da02b774`
`birdsnap`	16,000	25	0.36%	1.87%	`mteb/birdsnap` @ `fd230155`
`stanfordcars`	8,144	1	0.18%	1.34%	`mteb/StanfordCars` @ `09ffe9bc`
`renderedsst2`	6,920	1	0.16%	1.23%	`mteb/wds_renderedsst2` @ `c10537bc`
`stl10`	5,000	1	0.11%	1.05%	`mteb/stl10` @ `2da456fe`
`dtd`	3,760	1	0.08%	0.91%	`mteb/dtd` @ `96726183`
`oxfordpets`	3,680	1	0.08%	0.90%	`mteb/OxfordPets` @ `557b480f`
`caltech101`	3,030	1	0.07%	0.81%	`mteb/Caltech101` @ `011e51e5`

Caption fixes

MTEB's get_candidate_labels() targets zero-shot scoring, not training. Corrected here: an embedded \n in every label of the four file-backed tasks (Country211, GTSRB, ImageNet-1k, PatchCamelyon); no a/an agreement; cifar100 class 26 stored as cra (truncated crab); CLEVR closest-distance class 3 blank upstream, yielding the caption " shapes." for ~16% of rows; duplicate ImageNet strings (missile at 657 and 744, sunglasses at 836 and 837); gtsrb index 29 misspelled bicyle; Caltech's background_google clutter class dropped (102 -> 101); ucf101 CamelCase emitted verbatim; stanfordcars fully lower-cased; scimmir given its real per-figure captions instead of 5 generic prompts.

Shards are shuffled: samples are dealt at random across a rolling window of open shards, so the class-ordered sources (ImageNet, PatchCamelyon, CLEVR, Country211) do not produce single-class shards that defeat OpenCLIP's sample buffer.

Caveat

Every source except SciMMIR is also a MIEB zero-shot eval task. Training on these rows invalidates the corresponding MIEB score.

Downloads last month: 139

Total file size:

328 GB