Quickstart

End-to-end SAE pipeline on RFD3 in five commands.

Install


git clone https://github.com/RoseTTACommons/foundry && cd foundry
uv sync --extra sae
source .venv/bin/activate

Prepare data


# Download SafeProtein hazards + UniProt benigns
./data_pipelines/download_safeprotein.sh
./data_pipelines/download_uniprot_pdb_benigns.sh 500 100 300
 
# Merge and filter to ≤300 residues
uv run python -m data_pipelines.filter_pdbs \
  --sources tutorials/sae_data_rfd3_partial/sources.csv \
  --out tutorials/sae_data_rfd3_partial/sources_filtered.csv
 
uv run python -m data_pipelines.build_inputs \
  --sources tutorials/sae_data_rfd3_partial/sources_filtered.csv \
  --out tutorials/sae_data_rfd3_partial/train_inputs.json \
  --model rfd3 --partial-t 5.0

Collect activations


saffron collect model=rfd3 \
  hooks=rfd3_partial \
  inputs=tutorials/sae_data_rfd3_partial/train_inputs.json \
  out_dir=outputs/collect/train

Train SAE


saffron train \
  activations_path=outputs/collect/train/activations/activations.h5 \
  hook_name=block12 \
  steps=20000

Evaluate


saffron eval \
  checkpoint_path=outputs/train/matryoshka_batch_top_k/block12/<timestamp>/final.pt \
  activations_path=outputs/collect/train/activations/activations.h5 \
  hook_name=block12

For RF3 (sequence-only), see the RF3 pipeline. No PDBs required.