dnagpt/bio2nl-pretraining-2026-10

Bio2NL: pretraining materials

This is the dedicated pretraining companion to task data and the 313-checkpoint model bank. The GitHub project documents the experimental routes. This release covers the October 8 shared-BPE from-scratch EP/ES/EE experiment. It is not a copy of GPT-2's historical pretraining corpus or a release of every earlier pilot corpus.

This repository supplies the actual filtered English and protein pretraining text, discoverable dataset configurations, protein binary pools/streams, all four row indices and exposure ledgers, and a portable reconstruction helper. The October 17 addition publishes the two original English JSONL files without resampling or changing their text, membership or order. Existing corpora, token streams, provenance and model weights are unchanged.

What is included

Pool	Selected documents/sequences	Tokens in full pool, including EOS	Actual input stream	Distribution
English common	361,356	446,730,942	134,217,728	Actual filtered text + exact selection/index; regenerate token streams locally
English extra	361,643	447,059,806	134,217,728	Actual filtered text + exact selection/index; regenerate token streams locally
Natural protein	137,736	19,918,409	134,217,728	Raw sequences + full token pool + repeated stream
Shuffled protein	137,736	20,374,008	134,217,728	Raw sequences + full token pool + repeated stream

Protein raw pools each contain 53,988,068 amino-acid characters. Shuffling preserves each sequence's composition. Protein sources use the accepted Swiss-Prot 2026_03 pretraining train pool, with its recorded split and SCOPe exclusion policy, not the 5,000 short-query BLAST bank used to construct pair tasks. The protein pretraining pool rejects annotated fragments and non-canonical amino acids, deduplicates exact sequences, assigns whole operational 30%-identity components to train/validation/test, then removes detectable SCOPe domain matches (identity at least 30%, domain-query coverage at least 80%, E≤0.001) before retaining train entries. There is no 40–250-AA restriction on this pretraining pool. Detailed original source, split and exclusion receipts are in provenance/protein_*.json. English sources are exactly shards 0–7 of Skylion007/openwebtext at revision 79d93d786212f7344586290adb811d4ae6a1762c, followed by recorded filtering and deduplication; this is not the entire OpenWebText corpus.

indices/*.index.jsonl.gz identifies every retained record, original row ID, text hashes, token offset and token count. English indices contain no document text or reversible token IDs. exposure/ records every complete and partial presentation. provenance/ preserves original source receipts and filtering counts. Absolute paths in historical receipts identify original inputs; use this repository's relative paths with the reconstruction helper.

Conditions and tokenizer

• EP: common English + natural protein.
• ES: the same common English + shuffled protein.
• EE: common English + extra English; twice the English exposure of EP/ES.

Each condition has 268,435,456 input tokens, 16,384 optimizer updates and one pretraining seed (0). All three use a 12-layer, 12-head, 768-dimensional GPT-2-style model with 110,025,216 language-model parameters and 512-token context. Subsequent identical protein-pair classification fine-tuning (FT) uses seeds 1–20 for each condition. Twenty fine-tuning seeds are not twenty independent pretraining runs.

The new 32,000-entry byte-level BPE tokenizer was trained on 16 MiB English + 16 MiB natural protein. This tokenizer sample budget differs from model pretraining exposure. PAD=0, EOS=1, SEP=2; no UNK token. Full pools append EOS to each record. Binary files are little-endian uint16; streams are 262,144 blocks of 512 IDs. Protein streams repeat the pool before a final prefix; English streams take a single prefix without repetition. The training loader separately shuffles blocks; streams are not an epoch-by-epoch ordered log of optimizer inputs. See the original training code in the linked task release for batch scheduling.

Load the filtered corpora directly

from datasets import load_dataset
protein = load_dataset("dnagpt/bio2nl-pretraining-2026-10", "protein", split="train")
shuffled = load_dataset("dnagpt/bio2nl-pretraining-2026-10", "shuffled", split="train")
common = load_dataset("dnagpt/bio2nl-pretraining-2026-10", "english_common", split="train")
extra = load_dataset("dnagpt/bio2nl-pretraining-2026-10", "english_extra", split="train")

Pin the release commit in revision= for a reproducible run. Verify downloads against files_manifest.json.

Verify text and regenerate the original token streams

The two English files total 1,496,536,839 compressed bytes (about 1.50 GB). Each record includes its actual text, frozen source position, upstream shard identity and text hashes. english_text_release.json records exact file identities and counts. These are pretraining documents, not PAWS-X sentence pairs or the constructed English relatedness evaluation pairs. The latter remain separately documented in the task-data release.

The selected pools are larger than the text exposure used by the models: each English training stream takes the first 134,217,728 tokens of its pool, without repetition. Downloading the complete pool does not mean every document was presented to a model. The frozen exposure/ ledgers specify the actual complete and partial presentations.

Download this repository and install the reconstruction dependencies recorded in reconstruction_manifest.json (Python 3.12; no PyTorch or GPU required). The verification command checks every English record against the original selection index, source metadata and hashes; it does not retokenize the full pools.

python code/verify_english_text.py --root .
python code/reconstruct_pretraining.py --pool english_common --output ./rebuilt-common
python code/reconstruct_pretraining.py --pool english_extra --output ./rebuilt-extra
python code/reconstruct_pretraining.py --pool protein --output ./rebuilt-protein
python code/reconstruct_pretraining.py --pool shuffled --output ./rebuilt-shuffled

Use an empty output directory per reconstruction invocation. By default the helper verifies and reads the published corpus directly, then checks regenerated token pools and streams against their original SHA256 receipts. It also copies the original index and exposure ledgers unchanged. A damaged local corpus fails verification; it is never silently replaced. The helper restores frozen membership, without rerunning or changing the filtering rules. Whole-pool tokenization can take substantial CPU time; allow about 6 GiB for the downloaded materials and regenerated English pools/streams, with additional space for caches.

The original upstream route remains available as an independent reconstruction option:

python code/reconstruct_pretraining.py --pool english_common --from-upstream --upstream-cache ./owt-cache --output ./restored-common --download
python code/reconstruct_pretraining.py --pool english_extra --from-upstream --upstream-cache ./owt-cache --output ./restored-extra

That route verifies all eight pinned upstream shard identities, selects the frozen source positions, and checks the reconstructed text. Original gzip serialization uses Python's default compression, an empty stored filename and mtime=0; byte-identical reconstruction can depend on the original compression implementation. Allow about 10 GiB of working storage when retaining upstream shards as well. code/original_prepare_joint_data.py preserves the original builder and its workspace dependencies as provenance.

We acknowledge OpenWebText Corpus (Gokaslan et al., 2019). Its CC0 designation concerns dataset packaging; rights in the underlying website texts remain with their respective rights holders. Bio2NL claims no ownership of those texts and grants no additional blanket license for them. See LICENSES.md and the upstream attribution and notice policy. English token bins and the separate tokenizer-training sample are regenerated locally rather than duplicated here.

Publication checks cover exact corpus bytes, every selected English record and index, sample reconstruction/tokenization against frozen inputs, and the prior full protein token/stream regeneration. They do not claim a new full English retokenization or clean-machine model-training rerun.

File map

• corpora/: actual filtered English common/extra and natural/shuffled protein JSONL.
• tokenized_full/: complete protein/shuffled pools.
• streams/: exact protein/shuffled input streams.
• tokenizer/: the original tokenizer and training receipt.
• indices/, exposure/: all four pools' text-free selection and exposure records.
• provenance/, reconstruction_manifest.json, english_text_release.json: sources, original protocols, release additions and expected identities.
• code/: portable reconstruction plus immutable original builder.

Source-specific attribution is in LICENSES.md. Large transfer scores on the constructed English cosine proxy should not be described as human-paraphrase accuracy or evidence of a uniquely beneficial protein pretraining signal; the stronger EE control and all seeds are retained in the experiment release.