biomamba-780m

BioMamba-780m

Mamba2-780m continued-pretrained on PubMed/MEDLINE abstracts, mixed with Wikipedia and C4 to prevent general-domain forgetting.

This is the pretraining-only checkpoint. Downstream SFT variants are released separately.

Loading

The checkpoint is saved in mamba-ssm native format (not the transformers Mamba2Config format), so load it with mamba-ssm directly:

from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
from transformers import AutoTokenizer

model = MambaLMHeadModel.from_pretrained("zmzfpc/biomamba-780m", device="cuda", dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("zmzfpc/biomamba-780m")

AutoModelForCausalLM.from_pretrained will not work on this config.

Training data

• PubMed / MEDLINE abstracts: cyrilzakka/pubmed-medline (revision 432681e19469e93e6c42878d5f41fec400974fb8)
• Wikipedia: wikimedia/wikipedia, config 20231101.en (revision b04c8d1ceb2f5cd4588862100d08de323dccfbaa)
• C4: allenai/c4, config en (revision 1588ec454efa1a09f29cd18ddd04fe05fc8653a2)

Citation

Paper: BioMamba: Domain-Adaptive Biomedical Language Models (arXiv:2408.02600). Code: https://github.com/LeoYML/BioMamba

@article{yue2024biomamba,
  title   = {{BioMamba}: Domain-Adaptive Biomedical Language Models},
  author  = {Yue, Ling and Zhu, Mingzhi and Xing, Sixue and Pan, Shaowu and
             Chenthamarakshan, Vijil and Wang, Yanbo and Cao, Yunning and
             Das, Payel and Fu, Tianfan},
  journal = {arXiv preprint arXiv:2408.02600},
  year    = {2024}
}