Code and per-example results for the paper:
Scientific Data Composition as a Capability-Shaping Mechanism for Foundation Models:
Evidence from Biological Continued Pretraining — Liang Wang (HUST).
The study has two parts on a 26B-parameter Mixture-of-Experts model (Gemma-4-26B-A4B):
Part I — training-free re-analysis of one checkpoint lineage (instruction-tuned base →
biological CPT → SFT) across four capability axes.
Part II — a controlled seven-model experiment:… See the full description on the dataset page:
https://huggingface.co/datasets/dnagpt/OmniGene-4-bio.
