luban-agi / tool_use_benchmark

This is a tool-learning datasets for tool using!

English | 中文版

TLB

This is a tool-learning datasets.

Data Describetion

We have noticed that there are many high-quality datasets in the field of tool learning, such as Tool-Bench、Api-Bank、Tool-Alpaca, etc.

However, different tool learning datasets do not have a unified data format or evaluation standards for assessing the performance of large models on tool usage.

Consequently, we collected the aforementioned datasets and performed the following data processing tasks, including aligning the data formats, extracting parser keys, and constructing question-answer pairs.

Additionally, we have open-sourced a Chinese API-based tool dataset for research and learning purposes in tool learning.

Data Overview

DataseName	language	ToolRetrieval	Tool Filler	Tool Parser
ToolBench-g1	en	423734/484	-	-
ToolBench	en	423734/484	44939/484	19354/215
APIBank	en	429/79	429/79	429/79
ToolAlpaca	en	5911/198	5911/198	5803/198
ToolLuban	zh	10239/234	10239/234	5605/185

• Tool Retrieval：
  • data structure：<query, api_name, api_desc>
• Tool Filler：
  • data structure：<query, api_spec, api_param>
• Tool Parser：
  • data structure：<query, api_spec, response, select_keys

A/B：A/train_data_size + B/test_data_size

Metrics

• Tool Retrieval：
  • metric：Recall Top@1、Top@3、Top@5, to assess the retrieval effectiveness of the tool dataset
• Tool Filler
  • metric：Accuracy, to assess the ability of tool filler.
• Tool Parser
  • metric：Accuracy, to assess the ability of tool parser.

Citing & Authors

If you find this dataset helpful, feel free to cite:

@software{TLB,
  author = {luban},
  title = {Tool Learning Benchmark},
  year = {2023},
  url = {https://huggingface.co/toollearning/tool_benchmark},
}