xLAM Function-Calling (APIGen)

xLAM Function-Calling, also published under the name APIGen, is a dataset of 60,000 verifiable function-calling examples released in June 2024 by Salesforce AI Research. Each example pairs a natural language query with one or more structured function or API calls and their arguments, drawn from 3,673 executable APIs across 21 categories. It is the training data behind Salesforce's xLAM family of large action models. The examples were produced by the APIGen automated pipeline and passed through three verification stages (format checking, actual function execution and semantic verification). Salesforce reports that human evaluation of a 600-example sample found a correctness rate above 95 per cent. Roughly 33,659 entries were generated with DeepSeek-V2-Chat and the remainder with Mixtral-8x22B-Instruct.

Access is via Hugging Face, where the data ships as a single JSON file, but it is gated. You must be logged in and accept the terms before downloading, which is why we file it as limited rather than open. The accompanying paper (arXiv 2406.18518) and the APIGen project site document the pipeline and schema.

For RAG and AI work this is fine-tuning data, not a retrieval corpus. Use it to teach a model to read an instruction, select the right tool from a set and emit a correct, well-formed call with valid arguments. It is best for building single-turn and parallel function-calling capability rather than long multi-turn agent planning.

On licensing, verify carefully because the two licences differ. The dataset itself is released under CC BY 4.0, which permits commercial use provided you give attribution, and it is not share-alike. The trap is that the derived xLAM model weights carry a separate non-commercial licence (CC BY-NC 4.0), so the data being commercial-friendly does not make the models so.

It sits alongside Toucan-1.5M, our other agentic entry. Toucan is far larger and captures live multi-turn MCP tool trajectories, whereas xLAM is a tighter, heavily verified set of function-calling examples. Use xLAM for clean call formatting and Toucan for multi-step tool planning.