​

Tool Calling Example

This guide explains how to use the examples in examples/tool_calling_example/ for evaluating and training models for tool/function calling capabilities. These examples primarily use Hydra for configuration.

​

Overview

The examples/tool_calling_example/ directory contains scripts for:

1. Local Evaluation (local_eval.py): Evaluating a model’s ability to make tool calls against a dataset.
2. TRL GRPO Integration (trl_grpo_integration.py): Fine-tuning a model for tool calling using TRL (Transformer Reinforcement Learning) with Group Relative Policy Optimization (GRPO).

A sample dataset.jsonl is provided in the example directory. For tool calling tasks, each entry in the dataset typically includes:

• messages: A list of conversation messages.
• tools: A list of tool definitions available to the model.
• ground_truth: The expected assistant response, which might include tool calls (e.g., {"role": "assistant", "tool_calls": [...]}) or a direct content response.

​

Setup

1. Environment: Ensure your Python environment has reward-kit and its development dependencies installed:
   Copy

   # From the root of the repository
   pip install -e ".[dev]"
   

2. TRL Extras (for trl_grpo_integration.py):
   Copy

   pip install "reward-kit[trl]"
   

3. API Keys: If using models that require API keys (e.g., Fireworks AI models for local_eval.py if not using a local model, or for downloading a base model for TRL), ensure necessary keys like FIREWORKS_API_KEY are set.

​

1. Local Evaluation (local_eval.py)

This script performs local evaluation of a model’s tool calling.

​

Configuration

• Uses Hydra and is configured by examples/tool_calling_example/conf/local_eval_config.yaml.
• The default configuration points to examples/tool_calling_example/dataset.jsonl.
• The script itself likely contains defaults for the model and reward function, or expects them as CLI overrides.

​

How to Run

1. Activate your virtual environment:
   Copy

   source .venv/bin/activate
   

2. Execute from the repository root:
   Copy

   python examples/tool_calling_example/local_eval.py
   

​

Overriding Parameters

• Change dataset path:
  Copy

  python examples/tool_calling_example/local_eval.py dataset_file_path=path/to/your/tool_calling_dataset.jsonl
  

• Other parameters (e.g., model name, reward function parameters) would typically be added to local_eval_config.yaml or passed as CLI overrides if local_eval.py is structured to accept them via Hydra.

Outputs are saved to Hydra’s default output directory (configured in local_eval_config.yaml as ./outputs/local_eval_tool_calling/${now:%Y-%m-%d}/${now:%H-%M-%S}).

​

2. TRL GRPO Integration (trl_grpo_integration.py)

This script provides a scaffold for fine-tuning a model for tool calling using TRL GRPO. Note: The script defaults to using a MOCK model and tokenizer. Using a real model requires code modifications in trl_grpo_integration.py and potentially conf/trl_grpo_config.yaml.

​

Configuration

• Uses Hydra and is configured by examples/tool_calling_example/conf/trl_grpo_config.yaml.
• Default dataset_file_path: dataset.jsonl (assumed to be in examples/tool_calling_example/).
• Default model_name: Qwen/Qwen2-0.5B-Instruct.
• Includes various grpo training parameters.

​

How to Run (with Mock Model by default)

1. Activate your virtual environment:
   Copy

   source .venv/bin/activate
   

2. Execute from the repository root:
   Copy

   python examples/tool_calling_example/trl_grpo_integration.py
   

​

Overriding Parameters

• Change dataset path or training epochs:
  Copy

  python examples/tool_calling_example/trl_grpo_integration.py dataset_file_path=my_tool_train.jsonl grpo.num_train_epochs=1
  

​

Using a Real Model (Requires Code Changes)

1. Modify examples/tool_calling_example/trl_grpo_integration.py to load your desired Hugging Face model and tokenizer (remove or conditionalize the mock model parts).
2. Ensure the prompt formatting in the script is suitable for your chosen model.
3. Update conf/trl_grpo_config.yaml with the correct model_name and adjust training parameters.
4. Run the script. If you added a flag like use_mock_model_tokenizer in the script/config, you might run:
   Copy

   python examples/tool_calling_example/trl_grpo_integration.py +use_mock_model_tokenizer=false model_name=your-hf-model-name
   

Outputs are saved to Hydra’s default output directory (configured in trl_grpo_config.yaml as ./outputs/trl_grpo_tool_calling/${now:%Y-%m-%d}/${now:%H-%M-%S}).

For more general information on Hydra, see the Hydra Configuration for Examples guide.