dnagpt/bio2nl-data-2026-10

bio2nl data for the October revision

This new repository organizes the data underlying the October GPT-2 language/protein transfer experiments. It is separate from dnagpt/bio2nl, the earlier release. Dataset construction and model-to-data mapping are the primary purpose. The accompanying model bank is dnagpt/bio2nl-models-2026-10.

Dataset map

Dataset ID / family	Construction	Rows and role
blast_id*_random/length/composition_v1/v2	Swiss-Prot local BLAST, four positive identity bands × three screened negative strategies × two paired views	24 views, each 5,000 positive + 5,000 negative; all is the union of forward validation and test
sft_blast_id30_40/id40_60/id60_80/id80_100/mixed	V2 with length negatives; mixed uses balanced-cycle subsets of the four bands	Five source training sets of 10,000 each; no source held-out split
scope_remote	Same-superfamily/different-family positives and different-superfamily negatives; paired identity <25%, both coverage ≥60%	Train 13,732; validation 3,000; test 3,000 (not scored in current follow-up)
dms_function	Measured WT/single-mutant functional retention versus low outcome	Train 3,684; validation 272; CALM1 test retained but unscored
pawsx_clean	Earlier component-disjoint repaired English splits	Train 49,258; validation 543; test 1,998
pawsx_original	Original English validation/test	2,000 each; remote/DMS route
english_cos60_70/70_80/80_90/90_100	MiniLM cosine relatedness and eligible permuted negatives	Each validation 136/test 544/confirmation 1,000; text-free reconstruction references
pretraining_EP/ES/EE	Shared BPE; English+protein, English+shuffled protein, English+extra English	268,435,456 input tokens per condition; English exposure doubles in EE
standard_reference	Earlier rebuilt operational similarity dataset	Validation 20,276/test 20,854; separate from local BLAST

The exact names, paths and counts are in dataset_catalog.json. Do not sum overlapping views, splits or reused source groups as independent examples. V1 means the gap-free BLAST-aligned B fragment; V2 means the complete stored B entry. V1/V2 are representation views, not successive release versions.

Construction details that matter

The Swiss-Prot mother library has 230,094 length-eligible entries (40–250 AA). The common query bank has 5,000 records but 4,919 unique strings. Each final band has 5,000 positive pairs, not necessarily the same 5,000 distinct A sequences. Positives use local BLAST E≤1e-10, identity [30,40), [40,60), [60,80), [80,100]%, and no hard coverage cutoff. Negative screening rejects reported hits at E≤0.001 and self/equal/known-positive inputs. The accepted length rule is 30% relative to the original positive B in both views; the early 20% candidate was superseded. Composition matching is a prioritization, not a hard identity threshold or proof of nonhomology.

protocols/ records actual parameters, BLAST commands, the forward A-hash split and mixed-group row selection. Forward splitting isolates exact A strings and permits disclosed B overlap; it is not a family-disjoint split. original/ preserves the frozen positive/negative tables. Materialized data/ retains exact labels and original SFT row order. PAWS-X training rows add an empty-string foundation_component missing-value marker because this metadata was only recorded for the held-out splits; original input text, labels and ordering are unchanged. Remote SCOPe splits are superfamily-disjoint; functional variants use whole-assay/WT groups. These are three different biological relation definitions.

Usage and reconstruction

from datasets import load_dataset
ds = load_dataset('dnagpt/bio2nl-data-2026-10', 'sft_blast_id80_100')
print(ds['train'][0])

The HF loader exposes validation and test for each BLAST view. The complete 10,000-row all.jsonl.gz union remains available as a directly downloadable artifact; it is not an additional independent split. Pin a repository commit for reproducibility. code/verify_data.py checks every released data-file hash and row count. code/restore_english_proxy.py reconstructs all 6,720 English proxy pair rows from the separately acquired pinned OpenWebText shard(s), checking each endpoint's exact text SHA256. Shard filenames can be shard_00000.parquet, train-00000-of-00080.parquet or openwebtext-00000.parquet. Acquisition URLs and revision are in pretraining/source_lock.json.

python code/verify_data.py
python code/restore_english_proxy.py --shard-dir /path/to/pinned/shards --output /new/english_pairs

Original builder/training scripts are captured in code/original/ with their historical path bindings. These captures and source locks are reproducibility evidence; complete end-to-end retraining from a clean public-only environment is not claimed. The current portable checks cover materialized data, source reconstruction and model loading separately.

Models, results and interpretation

model_catalog.json connects every retained published checkpoint to a named source dataset, fine-tuning seed, pretraining seed, serialization/pooling regime and weight hash. results/ includes all prescribed seeds, raw and validation-oriented scores, selected-seed/initial/confirmation links and route budgets. Validation determines any label flip; held-out accuracy is not flipped to maximize test scores. The confirmation cohort was new originally and reused by later routes. High teacher-relatedness accuracy is not interchangeable with human paraphrase accuracy or proof of biological semantic transfer.

See LICENSES.md for source-specific attribution. This release includes protein sequences, PAWS-X, model/metric metadata and code. OpenWebText-derived English text and reversible corpus token bins are reconstructed separately. Private reviewer correspondence and manuscript drafts are not part of the release.