salma-remyx/spacellava-1.5-7b

SpaceLLaVA-7B

Run the colab above to analyze attention patterns for SpaceLLaVA-7B using this TransformerLens notebook

See how SpaceLLaVA-7B performs on Q-Spatial-Bench in the colab below

Visualize SpaceLLaVA-7B attention in the following colab

An experiment inspired by Linear Spatial World Models in Large Language Models

Check out these additional resources for mechanistic interpretability techniques compatible with LLaVA-1.5 based VLMs:

• https://github.com/nickjiang2378/vl-interp/blob/main/demos/internal_confidence.ipynb
• https://github.com/clemneo/llava-interp