mieb-train

CLIP training data built from the MIEB zero-shot image-classification datasets, with the text side re-written so the class labels are usable as captions.

4,434,196 training rows across 23 datasets.

column	notes
image	passed through from the source repo, not re-encoded
text	the caption — this is what CLIP trains on
label	source class id (-1 where the source has no class list)
class_name	normalized class name
dataset_name	source dataset key — filter on this to select a subset

dataset	rows	source
ucf101	1,786,096 train	mteb/ucf101 @ e0618988
imagenet1k	1,281,167 train	mteb/wds_imagenet1k @ b54c9af9
scimmir	498,279 train	mteb/SciMMIR @ 2a10b6b1
patchcamelyon	262,144 train	mteb/wds_vtab-pcam @ 6544d35f
sun397	76,127 train	mteb/sun397 @ c684cff3
food101	75,750 train	mteb/food101 @ 7cca05bf
clevr	63,000 train	mteb/wds_vtab-clevr_closest_object_distance @ d2777bb7
clevr_count	63,000 train	mteb/wds_vtab-clevr_count_all @ 2e31935f
mnist	60,000 train	mteb/mnist @ cf6afcba
cifar10	50,000 train	mteb/cifar10 @ 69a62dd1
cifar100	50,000 train	mteb/cifar100 @ ac5511f8
country211	31,650 train	mteb/wds_country211 @ c3875324
fer2013	28,709 train	mteb/wds_fer2013 @ d9cabab3
gtsrb	26,640 train	mteb/wds_gtsrb @ a8f4bb6c
resisc45	18,900 train	mteb/resisc45 @ a5b7c0d4
eurosat	16,200 train	mteb/eurosat-rgb @ da02b774
birdsnap	16,000 train	mteb/birdsnap @ fd230155
stanfordcars	8,144 train	mteb/StanfordCars @ 09ffe9bc
renderedsst2	6,920 train	mteb/wds_renderedsst2 @ c10537bc
stl10	5,000 train	mteb/stl10 @ 2da456fe
dtd	1,880 test / 3,760 train	mteb/dtd @ 96726183
oxfordpets	3,680 train	mteb/OxfordPets @ 557b480f
caltech101	3,030 train	mteb/Caltech101 @ 011e51e5

Why not use MTEB's own prompts

get_candidate_labels() is written for zero-shot scoring, not training, and carries defects that would go straight into the captions:

• An embedded \n in every label of the four tasks that read templates/*.txt — MTEB uses readlines() with no .strip() and each file ends in 0a: "a photo of tench\n.", "a photo showing the country of Andorra\n.", "a close up photo of a 'stop\n' traffic sign.". Affects Country211, GTSRB, ImageNet-1k, PatchCamelyon — 100% of their labels.
• No a/an agreement, ever: "a photo of a accordion", "a photo of a airplanes".
• cifar100 class 26 is stored as cra, a truncated crab (broken in upstream uoft-cs/cifar100 too).
• CLEVR closest-distance class 3 has a blank name upstream, so MTEB emits the caption " shapes." for ~17% of the split. Named mid-distance here.
• ImageNet-1k has duplicate strings — missile at both 657 and 744, sunglasses at both 836 and 837 — so two class ids shared one caption. Disambiguated by synset.
• gtsrb index 29 is misspelled bicyle.
• caltech101 ships 102 classes including background_google, the clutter class, which has no caption and is dropped here.
• ucf101 CamelCase is emitted verbatim ("a photo of ApplyEyeMakeup").
• stanfordcars labels are entirely lower-cased, destroying brands (am general, bmw, mclaren mp4-12c).
• scimmir is not a class dataset at all — it has a real per-figure caption column, used here instead of 5 generic class prompts.

Caveat: evaluation contamination

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

Loading

from datasets import load_dataset

ds = load_dataset("PumeTu/mieb-train", split="train")
ds[0]["text"]

# one source dataset
dtd = ds.filter(lambda r: r["dataset_name"] == "dtd")

Build tooling: normalization rules, per-dataset corrections and the WebDataset exporter live alongside this dataset.