Glaive Function Calling v2

Glaive Function Calling v2 is a synthetic supervised fine-tuning dataset for teaching language models to call tools, published on Hugging Face by Glaive AI. It contains roughly 113,000 examples (the current card reports 112,960 rows), each a multi-turn chat conversation in which an assistant is given a set of function definitions and must decide when to invoke one, how to fill its arguments, and how to use the returned result to answer the user. It also includes cases where the right move is to decline, teaching a model not to invent a tool call when no available function fits.

Access is simple. The data is hosted openly on Hugging Face in JSON and Parquet, downloadable with the datasets library or over HTTP, with no gating or credentials. Because the examples follow a consistent shape of a system prompt plus a conversation, many derivative versions exist that reformat it into ShareGPT, Llama chat and other schemas, so it drops straight into common fine-tuning pipelines.

It is used to build tool-using systems rather than to evaluate them. Fine-tune an open base model on these conversations and it learns the mechanics of function calling, argument formatting and grounding an answer in a tool result. It is best as an accessible, well-understood starting corpus for a first tool-calling model, and it remains one of the most downloaded open function-calling fine-tuning sets.

The licence is Apache 2.0, as tagged on the card, with no share-alike or attribution burden beyond the usual notice retention, so it is safe for commercial products. The main caveat is that the tools and results are synthetic and simpler than real agent traces. That is exactly where our sibling entry Toucan-1.5M differs: Toucan captures far larger, multi-step trajectories against live MCP servers, so treat Glaive as the lightweight entry point and Toucan as the heavier, more realistic follow-on.