dnagpt/OmniGene-4-bio

gemma4-bio

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: continue pretraining the same base under one fixed recipe, varying only the data mixture (0/5/20/50% biological share, plus a protein/DNA composition ablation), turning the observational claim into a causal one.
• Part III — a small-model, matched-compute causal probe (GPT-2, 3 seeds): asks the reverse question — does biological CPT feed anything back into natural-language understanding — in a regime with capability headroom the 26B model lacks.
• Part IV — robustness and mechanism: does the Part II pattern hold across a 10× model-scale range (2.3B–26B)? does explicit replay recover anything uniform mixing might have cost? does mixture ratio operate through the same router-reorganization mechanism as full CPT?

Key finding (Part I)

Biological CPT does not cause catastrophic forgetting; it lifts the model on axes unrelated to biology, and SFT then narrows it back — a consistent CPT-lifts / SFT-narrows division of labor.

axis	metric	Base (it)	BioCPT	BioCPT+SFT
General	MMLU (5-shot)	0.646	0.776	0.635
General	ARC-C acc (25-shot)	0.367	0.700	0.382
General	HellaSwag norm (10-shot)	0.488	0.855	0.486
General	TruthfulQA-mc2	0.533	0.445	0.562
Coding	MBPP pass@1 (3-shot)	0.332	0.630	0.348
Biology	BixBench-TF (MCC)	0.232	0.924	0.361
Reasoning	CoT chain length (tok)	108.8	64.4	125.8
Reasoning	backtracks / gen	0.370	0.006	0.396

Vocabulary expansion alone (it → it-bio) is free (< 0.4 pt on every general metric), so the gains are attributable to CPT, not the tokenizer.

Key finding (Part II — controlled mixture experiment)

Varying only the CPT data mixture (100M tokens each, one fixed recipe, from it-bio):

model	bio%	MMLU	GSM8K	homology-std (MCC)	homology-remote (MCC)	BixBench (MCC)
M0-base	0	0.802	0.873	0.228	−0.023	0.850
M1-bio5	5	0.801	0.867	0.270	−0.055	0.882
M2-bio20	20	0.800	0.867	0.217	−0.074	1.000
M3-bio50	50	0.796	0.861	0.375	0.088	1.000

Raising biological share to 50% preserves the general axis (MMLU −0.6 pt, GSM8K −1.2 pt) while the biology axis rises monotonically — the two do not trade off. A composition ablation at fixed 20% share shows DNA most improves remote homology while protein most improves BixBench: the type of scientific data selects which capability is amplified. Data mixture is a controllable, general-preserving lever on the capability profile.

Key finding (Part III — small-model matched-compute causal probe)

GPT-2 (124M): shared warmup (nl_base, 300M NL tokens), then two arms continued for an identical 200M tokens / 3,051 steps / optimizer schedule — only content differs (NL-only vs. 50/50 protein+DNA). Mean over 3 seeds:

family	task	NL-arm	Bio-arm	Δ (Bio−NL), mean±sd
Structural	PAWS-en (paraphrase detection)	0.498	0.529	+0.031 ± 0.009
Structural	Anagrams / cycle-letters / insertion / reversed (character-level)	0.000	0.000	floor at this scale (uninformative)
Knowledge	HellaSwag	0.291	0.285	−0.006 ± 0.000
Knowledge	LAMBADA (long-range discourse)	0.333	0.188	−0.146 ± 0.004

A real, reproducible transfer (all 3 seeds agree in sign) to the direct NL counterpart of protein-homology detection — but a genuine cost to long-range discourse, and the character-level battery is simply unmeasurable at 124M (both arms floor at 0, consistent with the original GPT-3 paper's own report of this floor below tens-of-billions of parameters). Reported as a partial, not uniform, dissociation.

Key finding (Part IV — robustness and mechanism)

Scale robustness. Repeating the Part II text-only-vs-bio50 contrast at two smaller dense Gemma-4 sizes (E2B 2.3B, E4B 4.5B) alongside the 26B MoE result:

model	MMLU (text→bio50)	BixBench MCC (text→bio50)
E2B (2.3B)	0.576 → 0.578 (+0.2 pt)	0.479 → 0.623 (+14.4 pt)
E4B (4.5B)	0.694 → 0.695 (+0.1 pt)	0.680 → 0.781 (+10.1 pt)
26B-A4B	0.802 → 0.796 (−0.6 pt)	0.850 → 1.000 (+15.0 pt)

General-preservation and the BixBench lift both hold across a 10× parameter range. Homology (harder, lower-sample-count) is noisier below 26B — likely a capability-floor effect on weaker dense models, reported honestly rather than smoothed over.

Replay control. At fixed 50% biological share, 100M tokens (26B), explicit end-of-training text replay (10%/20% of tokens) vs. the uniformly-shuffled baseline:

model	replay	MMLU	homology-std MCC	BixBench MCC
M3-bio50	0% (uniform)	0.796	0.375	1.000
R1	10% (tail)	0.800	0.318	1.000
R2	20% (tail)	0.795	0.266	0.873

The general axis is already flat without replay — there was nothing to recover. Replay does not help and mildly hurts the biology axis (fewer effective bio tokens as replay fraction rises). A clean negative result.

Router mechanism. Forward-hook analysis of all 30 MoE routers across every checkpoint: the full-CPT lineage (it-bio→BioCPT, 8.7B tokens) shows a large increase in biology-vs-language routing divergence (0.219→0.451), but the Phase-2 LoRA mixture sweep (M0→M3, 100M tokens) shows no routing dose-response (0.376–0.385, flat) despite a clear capability dose-response — capability shaping and routing reorganization are dissociable mechanisms that converge only at large (full-parameter) training budgets.

Checkpoints & data

• BioCPT (merged): dnagpt/OmniGene-4-CPT-v2-merged
• Data / per-example outputs: dnagpt/OmniGene-4-bio
• Base is the public Gemma-4-26B-A4B instruction-tuned model.

Repository layout

scripts/   evaluation harness (all reproduce on a single 96GB GPU, sm_120 grouped-mm guard)
  run_general.py         MMLU / ARC / HellaSwag / TruthfulQA (loglikelihood MC)
  run_coding.py          HumanEval / MBPP (generative + execution)
  run_bio_loglik.py      BixBench / homology (format-robust loglikelihood)
  run_cot.py             GSM8K chain-of-thought behavior diagnostics
  make_figs.py           paper figures
  run_phase1_*.sh        full sweeps over the 4 checkpoints
results/   per-(checkpoint, task) JSON + summary tables (PHASE1_*.md)
scripts_phase2/  controlled mixture-CPT experiment (Part II)
  prepare_pools.py       tokenize each source into a reusable pool
  run_cpt_mix.py         QLoRA CPT at a given data mixture (TAG, MIX, MIX_TOKENS)
  launch_parallel.sh     train the 7 models, one per GPU (auto-detects card count)
  bt2_load.py            reconstruct a CPT model (base + LoRA + trained embedding)
  run_eval.py / run_eval_bio.py   general + biology eval by model tag
  collect_results.py / make_figs.py   summary table + figures
  sm120_guard.py         dtype-safe MoE fallback for Blackwell (sm_120) training
results_phase2/  per-(model, task) JSON + phase2_summary.csv + figures
scripts_phase3/  small-model matched-compute causal probe (Part III)
  prepare_pools.py       tokenize NL/protein/DNA sources into GPT-2 pools
  train.py               warmup or matched-compute continuation (STAGE, MIX, TOKENS)
  run_tier1.sh            warmup -> 3 seeds x {NL,bio} arms, one GPU each
  run_eval.py / launch_eval.sh   NL-probe battery by model tag
  collect_results.py      per-seed delta + mean/sd, dissociation verdict
  *_local.yaml             custom lm-eval tasks for the GPT-3-paper character battery
                            (mirrors EleutherAI/unscramble, whose HF loading script is
                            no longer supported by current `datasets`)
results_phase3/  per-(model, task) JSON + tier1_summary.json
scripts_phase4/  robustness + mechanism follow-ups (Part IV)
  prepare_pools_gemma_small.py   tokenize sources w/ E2B/E4B native tokenizer
  run_scaling.py / run_scaling.sh   LoRA CPT on E2B/E4B (dense, no router)
  scaling_load.py                reconstruct an E2B/E4B CPT model for eval
  run_eval_scaling.py / run_eval_bio_scaling.py / launch_eval_scaling.sh
  run_replay.sh                  replay-control sweep (reuses run_cpt_mix.py
                                  with REPLAY_TAIL_FRAC/REPLAY_TEXT_SOURCE)
  run_eval_replay.sh              eval the replay-control models
  run_router_analysis.py / run_router_sweep.sh   30-router forward-hook JS
                                  divergence analysis across every checkpoint
  make_figs_part4.py              scaling / replay / router figures
results_phase4/  scaling + replay eval JSON, router/ = per-checkpoint routing reports
paper/     LaTeX source (Part I--IV sections), figures, compiled PDF

Reproducing

All benchmarks are public (MMLU, ARC, HellaSwag, TruthfulQA, MBPP, HumanEval, GSM8K, BixBench); no proprietary data. General/coding use lm-evaluation-harness 0.4.12 at community-standard few-shot. Example:

python scripts/run_general.py --ckpt <checkpoint> --tag mymodel \
  --tasks mmlu arc_challenge hellaswag truthfulqa_mc2 --out-dir results

Note: on Blackwell (sm_120) the scripts install a grouped-mm guard (transformers.integrations.moe._can_use_grouped_mm = lambda *a, **k: False) before any forward.

Citation

@article{wang2026biocpt,
  title  = {Scientific Data Composition as a Capability-Shaping Mechanism for
            Foundation Models: Evidence from Biological Continued Pretraining},
  author = {Wang, Liang},
  year   = {2026},
  note   = {preprint}
}