Work

Controlled-exposure checkpoints (E3) for resolution-aware sequence auditing

Year: 2026
Type: dataset
Abstract: Eight final training checkpoints from the controlled-exposure experiments (E3) behind the manuscript Resolution-aware sequence auditing for genomic variant-effect evaluation (Bo Han, Lanzhou University). Each run trains a 1.149 x 10^8 parameter genomic language model from scratch on the same prokaryotic base corpus for 61,065 steps (1,000,000,440 predicted tokens); the conditions differ only in which synthetic sequence blocks were injected into that corpus and at what dose. C0 is the unexposed baseline, C1 to C3 receive real homolog windows at x1, x10 and x100, C4 receives a first-order dinucleotide-preserving surrogate at x10, C4x1_order2 and C4x10_order2 receive a second-order empirical trinucleotide surrogate at x1 and x10, and C5 receives a second real x1 stream under an independent seed. Every file is a PyTorch checkpoint holding the state dict, the training configuration, the final step and the parameter count. The deposit also contains a forward-pass verification in which four fixed 1,000 bp natural homolog sequences were rescored on CPU in float32 against the saved per-item scores of the archived evaluations: all 32 item-model comparisons passed at a tolerance of 1e-4 nats per token, with a largest absolute difference of 3.37e-5. The sequences and the per-item values are stored in that file, and SHA256SUMS lists the digest of each checkpoint. The first-order surrogate at x1 (condition C4x1) was trained on a rented instance and its final checkpoint was not located when this inventory was assembled; Table S20 of the manuscript records that and the two substitutes used in its place. Licences: checkpoints CC-BY-4.0, analysis code MIT.
Source: ScienceDB
Author: Bo Han
Institution: Lanzhou University
Language: English
Open Access status: green
OpenAlex ID: W7221864668
Cites:
Cited by:
Related to: 10