Search

Few direct matches — filled in with the latest updates.

Tag: #tool-retrieval1 results

HumanMCP Dataset for Realistic MCP Tool Evaluation

HumanMCP Dataset for Realistic MCP Tool Evaluation

HumanMCP is the first large-scale dataset of diverse, human-like user queries for evaluating MCP servers' tool retrieval performance. It covers 2800 tools across 308 servers, built on MCP Zero, with multiple unique user personas per tool capturing precise to ambiguous intents. This addresses gaps in existing benchmarks lacking realistic user interactions.

ArXiv AIResearchMar 2#dataset#benchmark#tool-retrieval