mieb-train

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

Currently contains: dtd (mteb/dtd @ 96726183) — 3,760 train / 1,880 test.

Columns

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
class_name	normalized class name
dataset_name	source dataset key, so more can be appended later

Why not just use MTEB's prompts

MTEB's get_candidate_labels() is written for zero-shot scoring, not training, and carries defects that would land straight in the captions. Across the 23 MIEB datasets those include: an embedded \n in every label of the four file-backed tasks ("a photo of tench\n."), no a/an agreement, cifar100 class 26 stored as the truncated cra instead of crab, a blank class name for CLEVR closest-distance index 3, duplicate ImageNet strings (missile at 657 and 744), and gtsrb index 29 misspelled bicyle.

For DTD specifically the fix is narrow — snake/CamelCase is not an issue here, but the template must say "texture" (a photo of an interlaced texture.) and the article has to agree.

Caveat: evaluation contamination

DTD is also a MIEB zero-shot eval task, scored on its test split. The test split is included here and is clearly labelled — do not train on it if you intend to report DTDZeroShot.

Loading

from datasets import load_dataset
ds = load_dataset("PumeTu/mieb-train", split="train")
ds[0]["text"]   # 'a photo of a banded texture.'