dnagpt / laya-bio-models

Laya-Bio: final experiment checkpoints

Author: Liang Wang, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology.

This collection contains the 12 final checkpoints used in Laya-Bio: Candidate Scoring and Reliability on Short Biological Sequences : raw , full_bpe , b1 , and text_only , each at seeds 20260922, 20260923 and 20260924. The exact original Laya initialization is included under artifacts/laya_model/ , with its pinned upstream revision and attribution. Pilot, smoke and superseded experiment checkpoints are not the reported models.

Each checkpoint preserves its original model.safetensors , encoder/configuration, tokenizer and biological representation files. checkpoint_index.json lists paths, SHA-256, calibrated temperatures and tensor metadata. Every released final checkpoint file matches the frozen test manifest. Weights were not cast, quantized or retrained for this release.

• Benchmark data, predictions and manuscript
• Historical corpora and BPE training samples

Download

from huggingface_hub import snapshot_download

root = snapshot_download("dnagpt/laya-bio-models")
# Download only one model and its loading code when desired:
one = snapshot_download(
    "dnagpt/laya-bio-models",
    allow_patterns=["artifacts/laya_direct_legacy/full_bpe_seed20260922/**",
                    "scripts/**", "vendor/**", "checkpoint_index.json"],
)

This is a collection of custom Laya checkpoints, not one root-level Transformers model. Use the included pinned implementation and the loader in scripts/laya_locked_test.py (the run function shows the architecture and representation setup). Candidate models reload with laya_formal_experiment.fresh_reload ; the B1 fixed-class architecture uses laya_control_experiment.fresh_reload . Raw input uses the base tokenizer; biological-BPE input uses DirectBPE and its expanded tokenizer. Generic tokenizer text encoding does not replace the biological representation builder. The text-only control removes the sequence. Data and saved predictions live in the companion dataset repository; preserving the original directory paths allows the snapshots to be combined.

Dependencies are documented by vendor/laya/pyproject.toml and the experiment scripts (PyTorch, Transformers, tokenizers, safetensors, NumPy and SciPy). No model code executes merely by downloading files. Calibration temperatures are in the checkpoint index and training summaries; they were fitted on the calibration split.

Scope and reported behavior

The models were trained on two short-sequence closed-set tasks without additional neural continual pretraining. The historical BPE tokenizer was fitted externally; it is not a train-only tokenizer. Biological-BPE candidate scoring exceeds the implemented fixed-head control, while raw candidate accuracy is higher on both final test tasks. Candidate-order sensitivity and possible upstream exposure remain. These files support the reported comparisons, not claims of general biological reasoning or clinical validity.

All three seeds are provided without selecting a test winner. Test data have already been used for the reported results. Preserve the companion dataset's task-local label ordering. Historical corpora are a separate archive and their publication does not imply that these model checkpoints underwent continual pretraining on them.

See release_manifest.json and SHA256SUMS for the exact release inventory.