bio2nl — versioned data, code and aggregate results

Current audited release: 2026-10-03-v1

The release documentation provides the October 2026 protein data and exact code/aggregate package for the rebuilt 9-pretraining / 27-source-fine-tuning study.

• Code, aggregate results and Source Data: 190 explicit files, source capture commit b8d50d0ae61c8407ea8dcf62e47edef5969fc56c.
• Protein data: 8 actual Swiss-Prot-derived payloads plus manifest and attribution.
• Release checksums and UniProt attribution.

The release supports CPU verification, aggregate-report recomputation and figure reconstruction. A complete public-download-only raw-to-training pipeline is still incomplete. Model weights and tokenizer payloads are not included; the code package contains the identities of 36 local checkpoints. No new natural-language text, reversible NLP inputs, per-row predictions or private review materials are included in this release.

Licensing is per asset: the protein database-derived subset has the accompanying UniProt CC-BY-4.0 attribution; source files retain their stated notices. A repository-wide MIT label does not describe the new protein subset, and no new blanket license is assigned to historical code or omitted weights.

Earlier paths and files are preserved. The historical description below concerns earlier data/experiments and must not be treated as the construction or evidence for this release. The new study's small exploratory English-ranking difference does not establish semantic transfer or sufficient source-task learning.

Historical repository description (preserved)

bio2nl — data for Asymmetric Structural Transfer Between Natural Language and Biological Foundation Models

Datasets accompanying the bio2nl study of directional cross-domain structural transfer. Code: https://github.com/maris205/bio2nl

Contents

cpt_corpus/ — iso-token continued-pretraining corpora

Each file is exactly 50M GPT-2 tokens, differing only in sequence content (the iso-token ablation). Proteins are rendered as space-separated residues <protein> M K V ... </protein>.

file	content
cpt_protein.txt	real Swiss-Prot proteins (G2, main)
cpt_shuffled.txt	residue-shuffled proteins (G3; same AA freq/length, destroyed structure)
cpt_randomaa.txt	random amino acids from background frequency (G4)
cpt_text.txt	OpenWebText English (G1, control)
sprot.fasta	full Swiss-Prot source (575,503 proteins)
manifest.json	exact token counts per corpus

eval_nl/ — synthetic structural evaluation tasks

• dyck_L{40,60}_t3_*.jsonl — nested-bracket well-formedness (hard, non-saturating long-range structure)
• longrange_d{8,16,24}_*.jsonl — subject–verb agreement across distractors

Related

Protein-homology pairs use the existing dnagpt/biopaws release. PAWS-X, GLUE (CoLA/RTE), OpenWebText are downloaded from their standard sources.

Portable inputs, training entrypoints, and published models — 2026-10-03

The portable release adds public acquisition, exact current-study preparation, guarded source-training entrypoints, and standalone inference. Public downloads reproduced the accepted current-study inputs on CPU; no new full GPU training was run for this packaging check. The current PyTorch import/distribution mismatch is documented and rejected by normal GPU compilation.

All 9 pretrained and 27 source-selected classifiers include the exact shared tokenizer and loader. English documents and QQP examples are reconstructed locally, not added here. This is a current-study path, not a verified full historical raw-release rebuild. Asset-specific attribution and unassigned model licensing are stated in the release.