SKU AGT-4009 · Sold by External

ToolACE

Product specifications

SKUAGT-4009
Data typeAgent traces
Volume11,300 function-calling dialogues (multi-turn conversations)
Size on disk37.2 MB (data.json, 37,154,735 bytes)
FormatJSON (single data.json; each record = {system: string, conversations: list of {from, value}})
Access modelPUBLIC LICENSE
PricingFree · open dataset (HF)
Quality score
LicenseApache License 2.0
ToolACE is an open tool-learning / function-calling SFT dataset released on Hugging Face by the Team-ACE authors of the paper 'ToolACE: Winning the Points of LLM Function Calling' (arXiv:2409.00920). Each record is a system prompt that lists a set of available functions in JSON, followed by a multi-turn conversation whose assistant turns emit function calls in the format [func(param=value)] and whose tool turns return JSON results. The dialogues are produced by an agentic self-evolution synthesis process over a curated pool of 26,507 diverse APIs, with turns generated through interplay among multiple agents guided by a formalized thinking process. To ensure data accuracy, the authors apply a dual-layer verification system combining rule-based and model-based checks. The dataset is English/Chinese, released under Apache-2.0, and was used to fine-tune the ToolACE-8B model that scored on the Berkeley Function-Calling Leaderboard (BFCL). It is a supervised fine-tuning target: there is no live execution grader; correctness is established at synthesis time by the verification pipeline.