This repository contains pretrained encoders, cached features, and phase diagnostics data associated with the paper "When to Align, When to Predict: A Phase Diagram for Multimodal Learning".
The paper characterizes when contrastive (Cross-Alignment, CA) vs. predictive (Cross-Prediction, CP) self-supervised objectives recover shared signals in multimodal data, particularly for scientific domains.
Project… See the full description on the dataset page:
https://huggingface.co/datasets/Ilayk/mm_align_vs_pred.
