marin-community/mcp-atlas-easy
收藏资源简介:
MCP-Atlas-Easy是一个用于预训练(基础)语言模型的简单、单工具调用基准测试数据集,源自ScaleAI/MCP-Atlas数据集。它简化了多步骤工具编排任务,专注于最基本的技能:一个工具规范、一个明确无误的用户请求、一个正确的工具调用,然后停止。这使得它适用于没有聊天模板的基础模型的完成式评估。数据集包含220个唯一工具,每个工具对应一行数据,包括工具名称、服务器前缀、工具描述、输入模式(JSON模式)、用户请求场景、目标调用(JSON字符串)、就绪使用的完成提示和目标完成(JSON工具调用后跟EOS标记)。数据集的构建过程包括从MCP-Atlas提取工具信息、使用Claude Sonnet生成规范和场景,并进行程序化验证。注意事项包括部分工具的模式是推断的,场景和目标调用是合成的,数据集遵循CC-BY-4.0许可。
MCP-Atlas-Easy is an easy, single-tool-call benchmark for pretrained (base) language models, derived from ScaleAI/MCP-Atlas. It strips down multi-step tool orchestration to the simplest possible form of the same skill: one tool spec, one trivially unambiguous request, one correct tool call, then stop. This makes it usable as a completion-style eval for base models with no chat template. The dataset contains 220 unique tools, each represented as a row with fields including tool name, server prefix, tool description, input schema (JSON Schema), user request scenario, target call (JSON string), ready-to-use completion prompt, and target completion (JSON tool call followed by EOS token). Construction involved extracting tool information from MCP-Atlas, generating grounded specs and scenarios using Claude Sonnet, and programmatically validating each row. Caveats include that some tools have inferred schemas, scenarios and target calls are synthetic, and the dataset is released under CC-BY-4.0.




