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.

Complete paper FT and evaluation data (October 17)

For directly downloadable English target sentences, start with the English text guide. For the full experiment mapping, use the paper-to-data guide and machine-readable task catalog. The repository now exposes 46 loadable configurations: 41 with actual sequence/text pairs and five retained English-relatedness metadata configurations. The original payload files are preserved.

The additive completion includes the exact gap-cleaned Protein-PAWS Reference (14,000 archived train / 3,000 validation / 3,000 test), the identity-matched remote subset (1,696 / 1,832), and the 99,818 training pairs for the standard-reference feature baselines. These are distinct datasets. Existing ordinary/remote/functional protein FT sources and the exact 49,258-row English PAWS-X FT split are already included. All four original English-relatedness bands now have directly loadable metadata configurations, in addition to the expanded 2,000-pair target.

Every protein task input listed in the guide includes the actual sequences and labels. PAWS-X includes actual sentences. OpenWebText-derived English relatedness now also includes actual sentences in five new configurations without the _metadata suffix; the original source positions, hashes and reconstruction metadata remain available. The guide distinguishes training inputs, validation used for polarity, scored test sets, archived unscored splits and overlapping views. Mechanistic-analysis subset row IDs are provided separately.

October 17 extension: directly downloadable English relatedness sentences

The English text extension supplies the exact sentence pairs previously exposed only through reconstruction metadata. Use english_cos60_70, english_cos70_80, english_cos80_90, english_cos90_100, or english_cos90_100_expanded for actual sequence_a, sequence_b and integer label fields, together with unchanged row identities and source metadata. The five original _metadata configurations remain available. The earlier paper-task catalog describes its frozen 41-configuration release; the new extension catalog adds the five text configurations.

These are English evaluation targets for protein-trained classifiers. English-to-protein fine-tuning uses the actual PAWS-X sentence pairs in pawsx_clean/train; the constructed cosine targets were not the English fine-tuning source for those transfer experiments.

from datasets import load_dataset
pairs = load_dataset("dnagpt/bio2nl-data-2026-10", "english_cos90_100_expanded", split="test")
# Pin revision= to the commit containing this extension for exact reproduction.
# Each row includes sequence_a, sequence_b, label and source identities.

Construction keeps original OpenWebText sentences, applies the frozen sentence-quality and length filters, and scores fixed MiniLM sentence embeddings using cosine similarity. Positive bands are [0.60,0.70), [0.70,0.80), [0.80,0.90) and [0.90,1.00]. Within each split and band, negatives retain A and use a constrained one-to-one reassignment of the positive B sentences: cosine at most 0.30, GPT-2 token-length difference at most 30% relative to the original B, and no identical/conflicting or duplicate pairs. Sentence text is unchanged; A/B role frequencies are matched between labels. These are teacher-defined relatedness labels, with possible topic and lexical cues, rather than human paraphrase annotations. Pair selection did not use the evaluated classifiers' scores.

Each original band retains 136 validation, 544 test and 1,000 confirmation pairs. The expanded highest band provides validation136, original1000, new1000 and combined test2000 views; overlapping views are not independent datasets. Downloaded text can be used directly, without a new similarity search or upstream reconstruction. Original data, labels, splits, prediction inputs and results are unchanged. Source-specific attribution accompanies the text in LICENSES.md.

October 17 extension: 2,000 English relatedness pairs

The highest-cosine English extension adds 1,000 pairs to the unchanged 1,000-pair confirmation cohort. The combined set has 1,000 positives and 1,000 negatives. This is not an entirely new blind test. The original 544-pair discovery set remains in the original release, and the frozen 136-pair validation set is also provided as reconstruction metadata.

The new english_cos90_100_expanded_metadata configuration exposes validation (136), original1000 (1,000), new1000 (1,000), and test (the combined 2,000) as text-free metadata. These overlapping views must not be summed as independent examples. Row IDs, labels, pinned source positions and text hashes support optional reconstruction from the original OpenWebText shard. Actual English sentences are now provided by the english_cos90_100_expanded configuration in the text extension; the _metadata configuration is preserved for compatibility and provenance. The extension includes all retained model seeds, raw and original-validation-oriented results, same-seed comparisons and original-discovery-selected seed results. No model was retrained or reoriented for the expanded set.

Use the extension's reconstruction and verification instructions for the new target. The existing dataset_catalog.json and publication_status.json describe the original October 13 base release; the extension has its own data/results records. Existing base data and model files remain unchanged. The root files_manifest.json includes the October 17 extensions and the updated dataset card.

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; actual sentences plus retained metadata views
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	Train 99,818 (S14 feature baseline); validation 20,276/test 20,854; separate from local BLAST and Protein-PAWS Reference

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. For optional independent source verification, code/restore_english_proxy.py reconstructs all 6,720 original 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, actual OpenWebText-derived English relatedness pairs, model/metric metadata and code. Pretraining documents and tokenizer/stream reconstruction are provided separately in the pretraining companion. Private reviewer correspondence and manuscript drafts are not part of the release.