ToolBench

ToolBench is an open instruction-tuning dataset for tool use, built for the ToolLLM project by OpenBMB together with researchers at Tsinghua University, and presented as an ICLR 2024 spotlight. Released in 2023, it was among the first large-scale efforts to teach general tool-use ability to open language models rather than a fixed handful of functions.

The data is generated over 16,464 real-world REST APIs collected from RapidAPI Hub, spanning 3,451 tools across 49 categories. Instructions and their solution paths were produced automatically with ChatGPT, covering both single-tool and multi-tool tasks that require multi-step reasoning, and the release reports 126,486 instruction instances. Alongside the training data the project ships an evaluator (ToolEval) and the ToolLLaMA models fine-tuned on the corpus, so it works as a complete tool-learning pipeline rather than raw data alone.

You access it from the GitHub repository, with the dataset downloaded via Google Drive or Tsinghua Cloud as JSON. For RAG and AI work it is fine-tuning material, not a retrieval corpus: use it to teach a model to select tools, format API calls and chain several calls across a reasoning trace, then use ToolEval to score the result.

On licensing, take care. The repository carries an Apache 2.0 licence, but the maintainers state that ToolBench is intended solely for research and educational purposes, and the API content originates from RapidAPI and remains subject to RapidAPI's own terms of service. Treat commercial use as restricted and check those terms before shipping anything derived from the calls.

It is worth reading next to Toucan-1.5M, the other agentic entry we list. ToolBench is the earlier, foundational corpus built over RapidAPI REST endpoints, whereas Toucan is a newer and larger set generated against live Model Context Protocol servers. The two cover different tool ecosystems, and many practitioners now train on both.