Rumiii/Qwen-BioTool-1.5B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Rumiii/Qwen-BioTool-1.5B is a 1.5 billion parameter language model, fine-tuned from Qwen2.5-1.5B-Instruct, specifically designed for biomedical tool-calling. This model excels at selecting the correct tool and generating properly formatted function-call arguments given a biomedical question and available tool schemas. It is specialized for genomics, proteomics, and comparative biology tool-use across NCBI E-utilities, UniProt REST, and Ensembl REST API families.

Loading preview...

Qwen-BioTool-1.5B: Specialized Biomedical Tool-Calling Model

Rumiii/Qwen-BioTool-1.5B is a fine-tuned version of the Qwen2.5-1.5B-Instruct model, specifically optimized for biomedical tool-calling. It leverages QLoRA (4-bit) fine-tuning via Unsloth, with only 1.18% of its parameters being trainable, making it efficient for deployment.

Key Capabilities

  • Biomedical Tool Selection: Given a biomedical question and a set of tool schemas, the model accurately identifies the appropriate tool.
  • Function-Call Argument Generation: It generates correctly formatted arguments for the selected tool, facilitating automated API interactions.
  • Extensive Biomedical API Coverage: Trained on 127 tools across three major biomedical API families:
    • NCBI E-utilities: Including esearch, efetch, elink, and BLAST.
    • UniProt REST: For protein, proteome, and taxonomy lookups.
    • Ensembl REST: Covering coordinate mapping, variant effect prediction, and comparative genomics.

Training and Limitations

The model was trained on the gxx27/BioTool dataset, comprising 5,632 human-verified query-to-API-call pairs. While highly effective for its specialized domain, its tool coverage is limited to NCBI, UniProt, and Ensembl, and it does not cover clinical-facing APIs. The model's conversational ability is inherited from its base model and not reinforced during fine-tuning. Generated tool call arguments are JSON-encoded strings requiring an additional json.loads() step.