Which SE checkpoint was used to generate X_state for ST-SE-Tahoe?

[carloruggeri]

Hi Arc team,

I am using arcinstitute/ST-SE-Tahoe (from HF) for downstream inference/search experiments, and I am trying to determine which State Embedding checkpoint was used to generate the X_state embeddings used as input and target during ST-SE training.

From the public files, I can see that the few-shot ST-SE-Tahoe run uses:

• embed_key: X_state
• input_dim: 2058
• output_dim: 2058
• toml_config_path: /data/tahoe_se/generalization.toml

I checked config.yaml, version_0/hparams.yaml, var_dims.pkl, data_module.torch, and strings in checkpoints/final.ckpt, but I could not find the exact SE checkpoint/revision.

Could you clarify whether X_state for ST-SE-Tahoe was generated using:

1. arcinstitute/SE-600M, and if so which file/revision, e.g. model.safetensors, se600m_epoch16.ckpt, se600m_epoch4.ckpt; or
2. an internal/preprint SE checkpoint associated with Preprint-SE-167M-Human / SE-167M-Human; or
3. another checkpoint?

This matters because I want to generate compatible X_state embeddings for new AnnData files and compare ST-SE predicted embeddings to SE target embeddings in the same space.

Thanks!