csinva/tpr-fmri

Setup

• clone the repo and run pip install -e ., resulting in a package named tpr that can be imported
• download the linear encoding weights
  • OPT: download the weights here and move to the folder tpr-embeddings/fmri_voxel_data/llama_model/model_weights
    • rename the weights in that folder to wt_UTS01.jbl, wt_UTS01.jbl, wt_UTS03.jbl
  • LLaMA: download the weights here and move to the folder tpr-embeddings/fmri_voxel_data/llama_model/model_weights
    • rename the weights in that folder to wt_UTS01.jbl, wt_UTS01.jbl, wt_UTS03.jbl
• if everything is set up properly, you should be able to run the notebooks/01_module_example.ipynb notebook without any issues

Organization

• data: contains text and scripts for text to evaluate the models on
  • data/fmri: shows a sample test story of the type that the models were trained on
• voxel_data: contains metadata on the fMRI experiments
• tpr: contains main code for modeling (e.g. model architecture)
• notebooks: experiments in jupyter notebooks

Reference

This repo copies a lot of code from encoding-model-scaling-laws, which is the repo for the paper "Scaling laws for language encoding models in fMRI" (antonello, vaidya, & huth, 2023). See the cool results there! It also copies a lot of code from the repo for SASC.