rese1f/aurora

Employs token merging to reduce visual token count before LLM processing, enabling efficient video captioning while maintaining performance across inference and training. Includes the Video Detailed Caption (VDC) benchmark and integrates with lmms-eval for standardized evaluation, with planned support for HuggingFace transformers and SGLang deployment. Provides both XTuner and HuggingFace format weights, supporting continued training and inference through multiple framework backends.

139 stars. No commits in the last 6 months.