FineDNA post-training: grounded reasoning traces beat teacher distillation, 67% vs <60%

_lewtun · x · 2026-11-03

Lewis Tunstall (Hugging Face) shares how his team built the SFT/RL subsets of 🧬 FineDNA when no stronger teacher model exists.

Mid-training first: off-the-shelf Qwen3-14B produced learnable signal on only 17% of prompts; mid-training on 300B tokens of DNA + text raised it to 42%, growing the learnable pool from 952 to 7,200 prompts.
Skip the teacher entirely: they assembled training explanations from facts computed in code plus the model's own sense of how unusual a mutation looks — these "grounded reasoning traces" pushed accuracy to 67% via SFT, beating self-distillation (59.3%).
Self-distillation pitfalls: RLVR-then-SFT raised reward but the model hallucinated citations to nonexistent studies and databases.
Frontier teachers fail too: classifying 1,000 genetic variants, GPT-6.1 Sol gave only bare labels (encrypted reasoning), Gemini 4 Argon refused >50% of the time, Opus 5.5 ran out of tokens on 20%; open-weight models were barely better than random.

Takeaway: domain knowledge must enter the model via mid-training before RL has anything to learn from, and reasoning traces should be grounded in computable facts rather than black-box teacher outputs.

Original post →

FineDNA post-training: grounded reasoning traces beat teacher distillation, 67% vs <60%

More from Research