ToolACE is a synthetic dataset of function-calling dialogs for training large language models to call tools, produced by the automated pipeline of the same name described in ToolACE: Winning the Points of LLM Function Calling. It was built to give tool-learning data more diversity, complexity and accuracy than earlier tool-augmented training sets, and the paper uses it to fine-tune the ToolACE-8B model for Function calling.
Contents
Each record is a dialog among a user, an assistant and a tool, built around one or more APIs sampled from the dataset’s API pool. The pool holds 26,507 APIs across 390 domains, combining collected real APIs with synthesized ones whose definitions include nested parameter types such as lists of lists or lists of dictionaries. Dialogs cover four kinds of function-calling data: single calls, parallel calls, dependent calls and non-tool-use dialogs; the paper’s comparison table credits ToolACE, alone among the tool-augmented models it lists, with nested, parallel, dependent and multi-type data together. The paper’s appendix shows example records, including a parallel call made three times with different parameter values, a choice among five functions in one request, and cases in which the assistant declines because no available function fits or asks for missing required parameters.
Provenance
The APIs come from a self-evolution process that starts from API-related documents in LLM pretraining data — technical manuals, API documentation, product specifications, user guides and tutorials — builds a hierarchical API context tree of domains and functionalities, and then synthesizes and repeatedly mutates API definitions. Dialogs are generated by LLM-simulated user, assistant and tool agents, with the complexity of each dialog steered by the loss of the model being trained. Every sample passes a rule-based checker and model-based checkers, with the results overseen by human experts. The model and a subset of the data are published at https://huggingface.co/Team-ACE.
Use
ToolACE: Winning the Points of LLM Function Calling fine-tunes LLaMA-3.1-8B-Instruct on the data and evaluates it on Berkeley Function Calling Leaderboard (BFCL) and API-Bank. In an additional comparison at equal size (25,000 samples) on the same base model, the paper reports that training on ToolACE outperformed training on ToolLLM or xLAM data on BFCL, with 58.19 overall against 24.90 and 40.51.