[mieb] Fix siglip bug & add retrieval datasets

[gowitheflow-1998]

1. fixed SigLIP text tokenizer bug. padding=True->padding="max_length". Turns out this led to significant performance drop in zero-shot classification and retrieval that has texts (comparing the results here and https://github.com/embeddings-benchmark/tmp, e.g., MNISTZeroShot ~10 accuracy -> 80+ accuracy). Reference: https://github.com/huggingface/transformers/blob/v4.46.2/src/transformers/pipelines/zero_shot_image_classification.py#L147

2. Remove normalization in SigLIP single-modality encoding which constantly hurts linear probing performance (comparing CIFAR and MNIST results here and https://github.com/embeddings-benchmark/tmp).

3. Fix [mieb] google/siglip-large-patch16-384 fails on ImageNet10Clustering #1417

4. added EDIS and GLV-v2 I2I retrieval.

cc @Muennighoff

Checklist

• Run tests locally to make sure nothing is broken using make test.
• Run the formatter to format the code using make lint.

[isaac-chung]

Nice! I do wonder if points 1 and 2 should be applied to any other models 🤔

[gowitheflow-1998]

Nice! I do wonder if points 1 and 2 should be applied to any other models 🤔

looks like the tokenizer thing a siglip only thing! https://github.com/huggingface/transformers/blob/v4.46.2/src/transformers/pipelines/zero_shot_image_classification.py#L147

re normalization, looks like all other models by deafult do not normalize embeddings in single-modality encoding in our implementation, although we can potentially investigate normalizing them before fusion I think!