Skip to content

Latest commit

 

History

History
5 lines (4 loc) · 323 Bytes

README.md

File metadata and controls

5 lines (4 loc) · 323 Bytes

A small self-contained SAE trained on LLaVA-v1.5-7B, and can be seamlessly used to interprete it.

train_sae.py corresponding to the .py to train and evaluate SAE.

./llava/model/language_model/sae.py corresponding to the .py to SAE model structure.

Everything else remains the same as the original LLaVA.