Hugging Face open-sources FineDNA: full training data behind Carbon-X DNA model
_lewtun · x · 2026-11-02
Hugging Face's biology team released FineDNA, the complete open (Apache 2.0) training dataset behind Carbon-X — the model that taught Qwen3-14B to read DNA — covering all three training stages:
- Mid-training: 23 configs, 514.3B tokens (55% DNA + 45% biology text/instructions), 30.2B tokens consumed
- SFT: 205,923 conversations, 9,984 max token context
- RL: 98,433 prompts with code-verified reward functions, one GRPO epoch
Results: Carbon-X averages 66.7 across five DNA benchmarks, up 26.7 from the Qwen3-14B baseline of 40.0 (+19.3 from mid-training and SFT, +7.4 from RL). Among open models, only GPT-Sol 6.1 scores higher (67.5).
The card details data sources (eukaryote DNA, prokaryote mRNA, biology text, instructions), stage weightings, and loading recipes — 'keeping the bio frontier open.'
More from Models
- Fable 5.5 launch week teased: "see you real soon" — iruletheworldmo · 2026-11-02
- Amazon CEO expects just 6-8 AI models to be 'good enough' for most enterprise work — YvesMulkers · 2026-11-02
- ZenteiQ, backed by IndiaAI Mission, to unveil homegrown foundation models tomorrow — deepakns · 2026-11-02
- Altman: model that cracked Navier-Stokes is the next GPT, run as 10,000 agents for 4 days — Hesamation · 2026-11-02
- Tesla FSD Nears 17 Billion Miles, Adding 1.8M Miles Per Hour — XFreeze · 2026-11-02
- Gemini 4 Argon Fuel Pricing Doubles on December 1, Reddit User Cites Google's Own Pricing Page — Top_Departure_6901 · 2026-11-02