Datasets for building agents that use tools: function-calling and tool-calling examples, multi-step agent trajectories, and interactions with real APIs and Model Context Protocol (MCP) servers. The training and evaluation data behind systems that retrieve, reason, and act rather than simply answer.
10sources
Mercor's benchmark of 480 long-horizon, document-grounded professional tasks across investment banking, consulting and corporate law, with rubrics, gold outputs and an open execution harness for evaluating enterprise AI agents.
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.
The most-cited benchmark for general AI assistants: 466 real-world questions that need multi-step reasoning, live web browsing, file and image handling and tool use. Gated on Hugging Face, with test-set answers withheld for leaderboard scoring.
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.
NVIDIA's reinforcement learning dataset for conversational tool use, recasting expert multi-turn trajectories into 170,320 behaviour-cloning problems across 838 domains such as IT support, travel and EdTech. Released with the NeMo Gym framework for RLVR post-training.
The 500-task, human-validated coding-agent benchmark from Princeton and OpenAI. Real GitHub issues from Python projects where a model's patch must pass the project's test suite in a Docker harness, the de facto standard for issue-resolution agents.
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.
The largest open tool-agentic dataset: over 1.5 million trajectories synthesised from 495 real-world MCP servers spanning 2,000 plus tools, with multi-turn, sequential and parallel tool calls backed by real executions and error handling. Built by Agent-Ark and released under Apache 2.0, it is premier open data for training retrieval-and-tool (MCP) agents.
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.
Sierra's third-generation tool-agent benchmark. Its banking-knowledge domain puts document retrieval inside the agent loop, with a knowledge base of 698 policy documents the agent must navigate to act correctly. MIT-licensed, and still hosted in the tau2-bench repository.