Mathvista

[baberabb]

added mathvisata. Couple of issues:

1. The Qwen processor errors out if an image_dimension < 28 (defaults to do_resize=True and that fn requires a min). Added a padding function to handle the exception.
2. The mathvista evaluation code falls back to extracting the answer through GPT if the regex code fails. Not sure how we want to handle that.

currently scoring (0.624(ours) vs 0.632) for llava-hf/llava-onevision-qwen2-7b-ov-hf. For llava-hf/llava-onevision-qwen2-7b-ov-hf it's ~.30 vs ~0.58.

also added ai2d, which is a good candidate to test multiple-choice.