Awesome

MetaTool Benchmark: Deciding Whether to Use Tools and Which to Use

Introduction

We introduce MetaTool, a benchmark designed to evaluate whether LLMs have tool usage awareness and can correctly choose tools. It includes:

• ToolE Dataset: This dataset contains various types of user queries in the form of prompts that trigger LLMs to use tools, including both single-tool and multi-tool scenarios.
• Various Tasks: we set the tasks for both tool usage awareness and tool selection. We define four subtasks from different perspectives in tool selection, including tool selection with similar choices, tool selection in specific scenarios, tool selection with possible reliability issues, and multi-tool selection.
• Results on nine LLMs: We conduct experiments involving nine popular LLMs and find that the majority of them still struggle to effectively select tools, highlighting the existing gaps between LLMs and genuine intelligent agents.

ToolE Dataset

Dataset generation

We introduce the ToolE dataset with 21.1k diverse user queries related to tool usage. Each entry within the dataset comprises a user request (i.e., query) along with its corresponding tool name and tool description. These queries serve as triggers that prompt LLMs to utilize specific tools.

Dataset statistics

Dataset files

• Single-tool data: dataset/data/all_clean_data.csv
• Multi-tool data: dataset/data/multi_tool_query_golden.json
• All tool description: dataset/plugin_des.json
• meta data from OpenAI plugin store: dataset/plugin_info.json
• Merged data description: dataset/big_tool_des.json
• Embedding of tool description: dataset/tool_embedding.pkl
• Scenario tool list (Table 10 in the paper): dataset/scenario

Evaluation Results

Quick Start

First, create an .env file in the (put it next to src/generation/.example.env and include the same fields).

Now, you can run the following command for a quickstart (which downloads the model and prepares the data for you): bash quickstart.sh -m <model_name> -t <task>.

Alternatively, you can perform the below. Then, follow the results generation section.

Install the packages:

pip install --upgrade pip
pip install -r requirements.txt

Download the models:

• Set the HF_HOME environment variable in the src/generation/.env file to specify the Hugging Face model cache folder, e.g., HF_HOME=/path/to/your/huggingface/cache.

• --model-path: Specify the Hugging Face model repository name to download.

python src/generation/model_download.py --model_path lmsys/vicuna-7b-v1.3

Tool embedding

We use milvus to store tool embedding and conduct similarity searching.

To install and run milvus locally: https://milvus.io/docs/install_standalone-docker.md

Then run the following command to build a milvus database.

python src/embedding/milvus_database.py

Construct prompt data:

The pre-defined prompt templates are in src/prompt/prompt_template

If you want to generate the prompts for all tasks, run following command:

python src/prompt/prompt_construction.py

For single task prompts, run following command:

python src/prompt/prompt_construction.py [task]

Replace [task] with one of the following task options:

• similar: Similar tool selection.
• scenario: Scenario tool selection.
• reliable: Reliability tool selection.
• multi: Multi-tool prompt construction.
• all: All tasks.

Generate the results:

Parameters

You can generate results by running the run.sh script. You may need to modify the running parameters within the run.sh file to suit your specific needs.

• --test_type: Choose between tool_test_thought or tool_test_action depending on your testing needs.
• --model-path: Specify the Hugging Face model repository name.

sh src/generation/run.sh

Troubleshooting

If you face an import error from Python, you may need to add this directory to your Python path:

# Add sys path
src_path="$(pwd)/src"
export PYTHONPATH="$PYTHONPATH:$src_path"

Citation

@article{huang2023metatool,
  title   = {MetaTool Benchmark: Deciding Whether to Use Tools and Which to Use},
  author  = {Yue Huang and Jiawen Shi and Yuan Li and Chenrui Fan and Siyuan Wu and Qihui Zhang and Yixin Liu and Pan Zhou and Yao Wan and Neil Zhenqiang Gong and Lichao Sun},
  year    = {2023},
  journal = {arXiv preprint arXiv: 2310.03128}
}