Hugging Face dev drops FineDNA dataset teaching LLMs to chat with DNA
_lewtun · x · 2026-11-01
Hugging Face researcher Lewis Tunstall released FineDNA, a multi-stage dataset for teaching LLMs to understand DNA (announced with a "meaning of life" pun for Halloween):
- Mid-training: 300B+ tokens mixing genomic sequences, papers and web pages, bridging genomic LLM DNA representations with natural language semantics
- SFT: 200k samples on DNA computation, sequence analysis and variant effect prediction, grounded in real sequences with deterministic verifiers to prevent hallucination
- RLVR: 100k calibrated samples unlocking hard capabilities like localizing a DNA strand within huge sequences
The author claims models trained on it can "chat with DNA" and match frontier models on DNA understanding tasks. Dataset and code are open-sourced.
More from Fun
- mitsuhiko shares screenshot of AI agents politely apologizing to each other mid-work — mitsuhiko · 2026-11-01
- "Birdman got gaucho'd": eagle-masked figure in a beret becomes an AI-circle meme — StewartalsopIII · 2026-11-01
- AI-built Star Wars parody game SLOP EISLEY ships 0.6.70 update — jasonkneen · 2026-11-01
- Astra 6.2 oddly substitutes Russian letter 'п' for 'p' mid-sentence — teortaxesTex · 2026-11-01
- Can you name these three ML architectures from blank diagrams? — cneuralnetwork · 2026-11-01
- xAI's roon mocks Ed Zitron: this shitpost has more rigor than your last five Substacks — mjdramstead · 2026-11-01