TinyToolUse-Qwen2-0.5B-Calculator

A fine-tuned version of Qwen/Qwen2-0.5B trained for calculator tool usage using the Tiny Tool Use library.

Model Description

This model has been fine-tuned to understand and execute calculator tool calls. It can perform mathematical calculations by generating appropriate tool call syntax.

• Base Model: Qwen/Qwen2-0.5B
• Training Method: Supervised Fine-Tuning (SFT)
• Training Library: Tiny Tool Use
• Tool: Calculator (mathematical expressions)
• Parameters: 500M
• Architecture: 24 layers, 896 hidden dimensions

Training Details

• Training Data: 4 custom calculator examples
• Training Steps: 3 steps, 1 epoch
• Training Time: ~8.5 seconds
• Hardware: CPU training (RTX 4060 available but CUDA compatibility issues)
• Precision: float32

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("your-username/TinyToolUse-Qwen2-0.5B-Calculator")
model = AutoModelForCausalLM.from_pretrained("your-username/TinyToolUse-Qwen2-0.5B-Calculator")

# Example usage
prompt = "Human: What is 15 + 27?\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        inputs.input_ids,
        max_new_tokens=100,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Expected Output Format

The model should generate responses in the format:

tool_code: print(calculator(expression='15 + 27'))

Tool Definition

The model was trained with this calculator tool definition:

{
  "name": "calculator",
  "description": "Perform mathematical calculations",
  "type": "function",
  "parameters": {
    "type": "object",
    "properties": {
      "expression": {
        "type": "string",
        "description": "Mathematical expression to evaluate"
      }
    },
    "required": ["expression"]
  }
}

Training Examples

The model was trained on these examples:

• "What is 2 + 2?" → tool_code: print(calculator(expression='2 + 2'))
• "Calculate 10 * 5" → tool_code: print(calculator(expression='10 * 5'))
• "What is 100 / 25?" → tool_code: print(calculator(expression='100 / 25'))
• "Find the value of 3 to the power of 4." → tool_code: print(calculator(expression='3 ** 4'))

Limitations

• Small training dataset (4 examples)
• CPU-only training due to CUDA compatibility issues
• Limited to calculator tool only
• May require additional fine-tuning for production use

Framework

This model was trained using the Tiny Tool Use library, which provides:

• Multiple training methods (SFT, DPO, Teacher Mode)
• Flexible data generation strategies
• Integration with Berkeley Function-Calling Leaderboard
• Support for various LLM architectures

Citation

@misc{tinytooluse2024,
  title={Tiny Tool Use: Training Open-Source LLMs for Tool Usage},
  author={Bagel Labs},
  year={2024},
  url={https://github.com/bagel-org/bagel-RL}
}

License

MIT License - see the LICENSE file for details.