MixAtlas: Uncertainty-aware Data Mixture Optimization for Multimodal LLM Midtraining

Develop a data-mixture optimization tool for multimodal LLM midtraining. This platform should visualize how different data ratios affect specific benchmarks and allow users to optimize their mixture recipes dynamically.

Suggested repo: mix-atlas

"Stop guessing your data ratios: architect your multimodal training mixture with data-aware benchmarks."

Estimated effort: 100h