SigLip2 Math

This version of siglip2 is fine tuned on shiwk24/MathCanvas-Imagen using the code_derived_captions split. I trained for 1 epoch on 4M math images, with a random selection between the tikz code or caption using a batch size of 640.

This is not a classification model, since the loss function was pairwise contrastive loss. Use for embedding or downstream classifier training is recommended.