HuggingFace releases FineDNA dataset, turning Qwen3-14B into a DNA model that beats GPT-6.1
loubnabnl · reddit · 2026-11-01
HuggingFace's bio team open-sourced FineDNA, the full dataset and recipe used to turn Qwen3-14B into Carbon-X, the strongest open language model for DNA tasks:
- Midtraining: 300B+ tokens of genomes and DNA-rich text
- SFT: 200K conversations
- RL: 98K verifiable prompts
Code and training recipe are fully open and transferable to other LLMs. The accompanying benchmark chart shows Qwen3-14B starting at 40.0, rising to 59.3 with midtraining+SFT (+19.3) and 66.7 after RL (+7.4) — beating Opus 5.5 (60.9), GLM-5.3 (39.8), DeepSeek V4.1 Flash (47.1) and Kimi K3 (50.4), just shy of GPT-6.1 Sol's 67.5. The team also fixed a label leak in the BioReason benchmark.
Related event: HuggingFace Open-Sources FineDNA, 14B Model Tops DNA Benchmarks(2 posts)→
More from Models
- Robot Intelligence Index debuts: GPT-6 Astra tops robot model ranking at 1905 — haider1 · 2026-11-01
- Reddit rant: Google and OpenAI quietly nerfing $20-tier quotas by 30-50% — Outside-Clothes-8358 · 2026-11-01
- Pika launches H3 Flex: faster-than-real-time video model, up to 80% cheaper than H3Max — techNmak · 2026-11-01
- Prediction: Local Models Will Be the Biggest AI Story of 2027, Disrupting Inference Business — ZeYanjie · 2026-11-01
- Carbon-X: midtrained Qwen3-14B beats every open LLM on DNA benchmarks — LoubnaBenAllal1 · 2026-11-01
- DeepSeek-V4.1-Flash uncensored GGUF quant hits Hugging Face trending — AtomicChat · 2026-11-01