A ProSST adapter trained with ColabProSST.

This repository contains a SaprotHub/ColabProSST PEFT adapter and metadata for a ProSST downstream model. The task head is saved with the adapter, while the official ProSST backbone is loaded from AI4Protein/ProSST-2048.

ColabProSST uses amino-acid tokenizer input_ids together with ProSST structure token ss_input_ids. ColabProSST does not use SaProt AA+3Di merged tokens.

In ColabProSST, choose Hugging Face repository as the model source and enter TAVAT/ProSST-Protein-Regress-Test. The interface reads metadata.json, selects the matching official ProSST base model, and validates the task and structure vocabulary.

Use saprot/scripts/predict_prosst.py from SaprotHub to run prediction with this artifact outside the notebook.