Toucan-1.5M
Toucan-1.5M is an open dataset for training tool-using and agentic AI, released in October 2025 by Agent-Ark (The Agent Ark) alongside a technical report (arXiv 2510.01179). It is the largest fully synthetic tool-agent corpus published to date, with roughly 1.65 million trajectories generated from 495 real-world Model Context Protocol (MCP) servers spanning more than 2,000 distinct tools. The data was collected between June and September 2025.
What distinguishes Toucan is that trajectories were produced against live remote MCP servers rather than simulated tool responses. Queries were drafted by five models and quality-filtered, then trajectories were generated by three teacher models across two agent frameworks, with rule-based and model-based validation. The result covers multi-turn, multi-step, sequential and parallel tool calls, including real executions, tool outputs and genuine error handling, which is harder to fake than single-shot function-calling examples.
Access is straightforward. The dataset is hosted on Hugging Face in Parquet, split into subsets by teacher model, and the generation code is on GitHub (TheAgentArk/Toucan). For RAG and AI work it is primarily fine-tuning data, not a retrieval corpus: use it to teach a model to plan over tools, call APIs and recover from failures. The authors report that models tuned on Toucan beat much larger closed models on the BFCL V3 benchmark and extend the frontier on MCP-Universe.
The licence is Apache 2.0 with no territory or commercial restriction, so it is safe for commercial products, subject to the usual notice retention. The main caveat is operational rather than legal: the captured tools reflect the mid-2025 state of those MCP servers, and some community servers are flaky, so treat the tool surface as a snapshot. Unlike evaluation sets we list such as BEIR or CRAG, which measure a finished pipeline, Toucan is training material for the agent itself.
Related sources
Berkeley Function-Calling Leaderboard (BFCL)
The de facto benchmark for how well language models call functions, APIs and tools. Built by UC Berkeley's Gorilla project, it spans Python, Java, JavaScript and REST with simple, parallel, irrelevance-detection, multi-turn and agentic cases. Apache 2.0 and freely available.
Glaive Function Calling v2
A widely used open dataset of about 113,000 synthetic multi-turn chat conversations that include function calls and their results, made by Glaive AI. One of the most downloaded open datasets for fine-tuning models to call tools, released under Apache 2.0.
ToolBench
An open instruction-tuning dataset for teaching general tool use to language models, built for the ToolLLM project over 16,464 real-world REST APIs from RapidAPI across 49 categories, with single-tool and multi-tool, multi-step solution paths.
xLAM Function-Calling (APIGen)
Salesforce's APIGen-generated dataset of 60,000 verified function-calling examples spanning 3,673 executable APIs, each checked through format, execution and semantic stages. The training data behind the xLAM action models and core capability data for tool-using agents.