keerthanaSubru57/mcp-bench
收藏资源简介:
MCP-Bench任务数据集(v0.1)是一个包含70个任务的基准测试数据集,用于支持MCP-Bench项目——一个基于Anthropic的模型上下文协议(MCP)评估开源大型语言模型(LLM)代理可靠性的基准测试。数据集覆盖了多种MCP服务器类型,包括fetch(13个任务)、filesystem(14个任务)、github(10个任务)、memory(10个任务)、postgres(10个任务)和sqlite(13个任务)。任务按难度分为简单(24个)、中等(34个)和困难(12个),并按技能分为选择(40个)、组合(23个)和恢复(7个)。每个任务以JSON对象形式存储,包含唯一ID、描述、用户提示、所需服务器、设置操作、检查操作、标签、难度和技能等字段。数据集通过确定的检查操作符(如文件存在检查、内容匹配、SQL查询结果验证等)实现可复现的通过/失败判定,适用于工具使用、代理和基准测试研究。
The 70-task dataset behind MCP-Bench — a reliability benchmark for open-source LLM agents using Anthropics Model Context Protocol (MCP). It covers multiple MCP server types including fetch (13 tasks), filesystem (14 tasks), github (10 tasks), memory (10 tasks), postgres (10 tasks), and sqlite (13 tasks). Tasks are categorized by difficulty: easy (24), medium (34), and hard (12), and by skill: selection (40), composition (23), and recovery (7). Each task is structured as a JSON object with fields such as unique ID, description, user prompt, required servers, setup operators, check operators, tags, difficulty, and skill. The dataset uses deterministic check operators (e.g., file existence checks, content matching, SQL query validation) for reproducible pass/fail decisions, suitable for tool-use, agent, and benchmarking research.




