A research project exploring supervised fine-tuning (SFT) for biological tool calling using open-source large language models.

The primary objective is to build a language model that can recognize when external biological tools are required, generate structured tool calls, execute those tools, and return accurate natural-language responses.

• Learn the complete supervised fine-tuning (SFT) pipeline.
• Understand parameter-efficient fine-tuning (LoRA / PEFT).
• Build a production-style tool-calling agent from scratch.
• Explore biological AI applications through practical implementations.
• Develop a reusable framework for adding new biological tools.

• Supervised fine-tuning using LoRA.
• JSON-based tool calling.
• Automatic tool dispatching.
• Natural-language response generation from tool outputs.
• Dataset generation using GPT-4o.
• Versioned datasets and checkpoints.

Current tools include:

• gene_lookup
• search_pubmed
• classify_sequence
• final_answer

Additional tools will be added in future dataset releases.

bio-agent/
│
├── datasets/
│   └── v0.0.1/
│       ├── README.md
│       ├── metadata.json
│       ├── train.jsonl
│       └── test.jsonl
│
├── checkpoints/
│   └── qwen2.5-0.5b/
│       └── v0.0.1/
│
├── scripts/
│   ├── generate_dataset.py
│   ├── train.py
│   ├── evaluate.py
│   └── inference.py
│
├── tools/
│
└── README.md

User
  │
  ▼
Fine-tuned LLM
  │
  ▼
JSON Tool Call
  │
  ▼
Python Dispatcher
  │
  ▼
Biological Tool
  │
  ▼
Structured Result
  │
  ▼
LLM Summarization
  │
  ▼
Final Answer

Datasets are versioned independently from model checkpoints.

Each dataset version contains:

• Training data
• Test data
• Metadata
• Documentation

Example:

datasets/
    v0.0.1/
    v0.1.0/
    v1.0.0/

Each trained model is stored separately.

Example:

checkpoints/
    qwen2.5-0.5b/
        v0.0.1/
        v0.1.0/

    qwen2.5-3b/
        v0.1.0/

Every checkpoint includes the adapter, merged model (if applicable), tokenizer, and metadata describing the training configuration.

• Increase dataset size
• Improve prompt diversity
• Human quality assurance
• Add multi-tool examples
• Expand biological coverage

Planned additions include:

• Protein lookup
• BLAST sequence search
• Protein structure prediction
• Variant lookup
• Pathway lookup
• GO annotation
• Disease association
• Sequence alignment

• Larger Qwen models
• Multi-tool agents
• Tool planning
• Multi-step reasoning
• Reinforcement learning experiments

This repository serves both as a learning project and as a foundation for future research in biology-focused AI agents. Rather than relying on general-purpose assistants, the long-term goal is to build specialized systems capable of interacting with domain-specific biological resources through reliable tool use.

This project is intended for research and educational purposes.