Multi-Modal Foundation · PharMolix / OpenBioMed
BioMedGPT-R1
A biomedical language model providing chain-of-thought reasoning over text, small molecules, and protein sequences.
What BioMedGPT-R1 is
BioMedGPT-R1 is a multimodal biomedical language model developed by PharMolix and OpenBioMed. Designed for interactive dialogue and molecular reasoning, it processes queries combining natural language, small molecules, and protein sequences. The system produces answers alongside step-by-step reasoning traces, allowing researchers to evaluate the logical pathway behind each conclusion.
The model sits within the OpenBioMed ecosystem, which bridges text with biochemical data modalities. By generating explicit rationale rather than returning unverified answers, it addresses the need for interpretable deductions in computational biology and chemistry workflows.
How it works
The core language model component was fine-tuned from Llama2-7B-Chat using over 26 billion tokens extracted from millions of scientific articles in the S2ORC corpus, identified via PubMed and PubMed Central criteria. Fine-tuning used a context length of 2,048 tokens, a batch size of 192, a learning rate of 2e-5, and ran across five training epochs.
Integrated into a multimodal architecture, the model connects text representations with chemical and biological structures, generating chain-of-thought deductions across small-molecule and protein-sequence contexts.
Key facts
- Base model
- Llama2-7B-Chat
- Modalities
- Text, small molecule, protein sequence
- Context length
- 2,048 tokens
- Training data volume
- Over 26 billion tokens from S2ORC
- GPU memory requirement
- 40 GB
What you can use it for
- Auditable Molecular Reasoning
- Analyse chemical queries involving small molecules while inspecting the intermediate reasoning trace generated by the model.
- Protein-Sequence-Conditioned Dialogue
- Query the model on biological tasks that combine protein sequence information with natural language questions.
- Biomedical Literature Question Answering
- Address complex biomedical queries requiring domain-specific knowledge derived from PubMed Central and PubMed literature.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on Managed endpoint · text-generation
Runs on managed compute. Availability and startup time depend on current capacity.Run this model
BioMedGPT-R1
Strengths and limits
Good at
- Outputs an explicit reasoning trace to help researchers audit how intermediate conclusions are derived.
- Operates across multiple scientific modalities, including natural language, small-molecule data, and protein sequences.
- Trained on more than 26 billion tokens of domain-specific literature drawn from PubMed and PubMed Central articles.
Watch out for
- Requires 40 GB of GPU memory to run.
- Restricted to an input context length of 2,048 tokens.
- Must not be used to deliver services to the general public.
Inputs and outputs
You provide
Question
You get back
Response with reasoning trace.
Licence
Commercial use permitted · Apache-2.0
The Apache-2.0 licence allows use in commercial research and products.
Questions
Can I use BioMedGPT-R1 commercially?
Commercial use permitted. The Apache-2.0 licence allows use in commercial research and products.
How do I run BioMedGPT-R1 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 12s, and a run costs about 84 credits. Visitors without an account get 1,000 credits a month.
Which data modalities does BioMedGPT-R1 accept?
The model supports text prompts, small molecules, and protein sequences.
What distinguishes BioMedGPT-R1 outputs from standard language model responses?
It provides a chain-of-thought reasoning trace alongside its final answer, detailing the deduction steps taken to answer the prompt.
What training corpus was used for the underlying language model?
It was fine-tuned on over 26 billion tokens drawn from millions of biomedical publications in the S2ORC corpus, selected using PubMed ID and PubMed Central ID criteria.
What hardware is necessary to run the model?
The model requires 40 GB of GPU memory.
How to cite it
BioMedGPT: Open Multimodal Generative Pre-trained Transformer for BioMedicine, arXiv:2308.09442v2
Written from BioMedGPT-R1's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.