Mistral-BioMed-Tool-Caller-7B

A fine-tuned version of Mistral-7B-Instruct-v0.3 specialized for biomedical tool-calling. Given a biomedical question and a set of available tool schemas, the model selects the correct tool and generates properly formatted function-call arguments in Mistral's native [TOOL_CALLS] format.

Model Details

Property	Value
Base model	mistralai/Mistral-7B-Instruct-v0.3
Fine-tuning method	QLoRA (4-bit) via Unsloth
LoRA rank / alpha	32 / 32
Target modules	q/k/v/o_proj, gate/up/down_proj
Trainable parameters	83,886,080 (1.14% of total)
Training dataset	gxx27/BioTool (5,408 examples after filtering) + 540 general-instruction examples (~10% mix)
Epochs	1
Final training loss	~0.27
Hardware	Single Kaggle Tesla T4

Benchmarking

The model achieved 99.8% tool selection accuracy, 90.4% parameter name accuracy, 65.1% parameter value accuracy, 53.1% whole-call correctness, and 34.7% exact reference call matching.

Tool selection accuracy reached 100% across NCBI, Ensembl, and UniProt. The results demonstrate strong tool identification capabilities, while parameter generation and end-to-end call correctness remain areas for improvement.

The model was fine-tuned on the BioTool dataset. Further experimentation is needed to investigate the effects of training strategy and duration on overall tool-calling performance.

Coverage

Trained on tool-calling patterns across 127 tools spanning three biomedical API families:

NCBI E-utilities (esearch, efetch, elink, BLAST, and related endpoints)
UniProt REST (protein, proteome, and taxonomy lookups)
Ensembl REST (coordinate mapping, variant effect prediction, comparative genomics)

This covers genomics, proteomics, and comparative biology tool-use. It does not cover clinical-facing APIs such as ICD-10 lookup, drug databases, or clinical trial registries.

A small slice (~10%) of general-instruction data (HuggingFaceH4/no_robots) was interleaved during training specifically to preserve the base model's general conversational ability alongside its new biomedical specialization.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Rumiii/Mistral-BioMed-Tool-Caller-7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

tools = [
    {
        "type": "function",
        "function": {
            "name": "esearch",
            "description": "Search an NCBI Entrez database and return UIDs matching a text query.",
            "parameters": {
                "type": "object",
                "properties": {
                    "db": {"type": "string", "description": "Entrez database name"},
                    "term": {"type": "string", "description": "Search query"},
                },
                "required": ["db", "term"],
            },
        },
    }
]

messages = [{"role": "user", "content": "Search PubMed for articles on BRCA1 mutations."}]

inputs = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Training Data

BioTool: a biomedical function-calling dataset of 7,040 human-verified query-to-API-call pairs across NCBI, UniProt, and Ensembl.

@misc{gao2026biotoolcomprehensivetoolcallingdataset,
      title={BioTool: A Comprehensive Tool-Calling Dataset for Enhancing Biomedical Capabilities of Large Language Models},
      author={Xin Gao and Ruiyi Zhang and Meixi Du and Peijia Qin and Pengtao Xie},
      year={2026},
      eprint={2605.05758},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.05758},
}

This model was trained with Unsloth and Hugging Face's TRL library.

License

Apache 2.0

Rumiii/Mistral-BioMed-Tool-Caller-7B