tribev2-brain-encoding — Claude Code skill: fMRI prediction from video

What tribev2-brain-encoding does

TRIBE v2 comes from Meta and merges three encoders: LLaMA 3.2 handles text, V-JEPA2 handles video, Wav2Vec-BERT handles audio. A shared Transformer maps those representations onto the fsaverage5 cortical surface, roughly 20k vertices. The skill lets your agent run inference and training against it.

Inference needs only pip install -e .. Brain plots pull in the [plotting] extras with PyVista and Nilearn; full training pulls [training] with PyTorch Lightning and W&B. Weights come from HuggingFace via TribeModel.from_pretrained("facebook/tribev2", cache_folder="./cache"). You then build an events dataframe with get_events_dataframe and call predict.

Training from scratch runs into LLaMA 3.2 gating. You log in with huggingface-cli login using a read token and wait for access approval on the model page. Grid searches are written for Slurm and read DATAPATH, SAVEPATH and SLURM_PARTITION from the environment. No GPU means no training.