SarahDaakour/dark-whiteGPLM-data

dark-whiteGPLM Training Data

Character-level protein-sequence datasets used to train five GPT-2 Small models. See the model repo at https://huggingface.co/SarahDaakour/dark-whiteGPLM and the code at https://github.com/SarahD4/dark-whiteGPLM for full details.

Contents

• char_white/ — annotated proteins from photosynthetic organisms
• char_algae/ — algal proteome (400 k sequence subsample)
• char_dark/ — uncharacterised / "dark" proteins
• char_random_algae_biased/ — composition-matched random sequences (algae frequencies)
• char_random_full/ — max-entropy random sequences (uniform 20 AA)

Each directory contains input.txt (source sequences), train.bin and val.bin (tokenised uint16 arrays, 90/10 split), and optionally the source FASTA.