kevinlu4588/prot2text-v3-data

Prot2Text-V2 data with ESMC-6B features

Data for reproducing Prot2Text-V2 (arXiv 2505.11194) on LLMFold.

• raw/ the source parquet splits
• splits/ every row (248,315 train / 4,172 validation / 4,203 test), sequences truncated to 1024 residues; sequence_sha256 is the hash of the truncated sequence and keys the features
• esmc6b/ frozen ESMC-6B final-layer per-residue features, unbatched, 129 safetensors shards with .json manifests
• afdb/validation/ AlphaFold DB models for the first 400 validation rows, for fold scoring. path in index.json is absolute on the original machine; resolve it by file name.

Put esmc6b, splits and afdb under data/prot2text_v3/ of the LLMFold repo. Upstream sources keep their own terms (UniProt, AlphaFold DB, ESMC-6B).