huggingface/carbon

在 1T DNA token 上预训练的基因组基础模型，附带一套统一的零样本 DNA 评测基准，解决现有评测分散、难以复现的问题。 Hugging Face 的基因组语言模型，有 500M、3B、8B 三个版本，3B 是旗舰，效果追平甚至超过 Evo2 7B。训练数据 1T token（约 6T 碱基对），真核占大头，原核 10%，方便后续对特定物种做持续预训练。

💡 Takeaway: CARBON is not only memorizing local k-mer patterns. Its embeddings organize held-out DNA sequences according to biological structure, from genome-level taxonomic signal to local…

biology just got its tokenization war. Hugging Face released Carbon, a family of open generative DNA foundation models. the funny part is that “DNA is language” was not enough. BPE, the obvious LLM…

Carbon: Decoding the Language of Life

We also release new evaluation probes, including sequence-level perturbation tasks and Genomic-NIAH: a long-context retrieval benchmark built from real genomic sequences. The goal is to make DNA…