Understanding Information Storage and Transfer in Multimodal Language Models

This is the implementation of the interpretability and model editing experiments from NeurIPS 2024 paper : https://arxiv.org/abs/2406.04236.

================================================

Constraint Annotations

The constraint annotations are in ./data_constraints. The directory contains constraints for OK-VQA, Multimodal Known and Multimodal Movies.

================================================

Images used for Probe Dataset

For OK-VQA, we use the val set images from https://okvqa.allenai.org/. For Multimodal Known the images are at: Link 1 and for Multimodal Movies the images are at: Link 2.

================================================

Running the Scripts

Our codebase is built on Llava's code. Clone Llava and transfer the code from this repository to ./llava/eval .

1. Multimodal Causal Trace: python -m llava.eval.multimodal_trace --trace_answer <Answer for the prompt> --trace_question <Question> --trace_image <Image Path> --trace_constraint <Constraint in the question>

2. Multimodal Edit: python -m llava.eval.multimodal_edit --edit_prompt <Prompt used for running model editing> --edit_constraint <Constraints in th prompt> --edit_target <Target Answer> --edit_og_answer <Original Answer to the prompt>

================================================

Notes : Although the current script is optimized for Llava, these scripts can be modified towards applying it on any multimodal language model. Reach out to [email protected] for any questions.