【Agent 微调的核心秘密:Function Calling 和 MCP 数据怎么准备?】从 Agent 日志积累 badcase 后想微调大模型,数据格式是关键:1. Function Calling 格式一条好数据 = System(工具描述) → User(问题) → Assistant(tool_calls+参数) → Tool(结果) → Assistant(回答)注意 tool_calls 必须是严格 JSON,参数名要与 schema 一致2. MCP 格式更难MCP 靠 Prompt 约束输出,模型需学会输出标准文本标记(如 XML 标签),数据量建议 100 到 200 条起步3. 四种 badcase 数据策略- 工具选错:同一场景准备 5 到 10 种问法覆盖- 参数填错:边界值加标准化训练- 格式不听话:高强度反复训练(50 到 100 条)- 多步推理乱序:展示完整轨迹链4. 最佳实践- 一条数据展示完整多轮 tool_call 轨迹,比零散截取效果好- 混合 20% 到 30% 通用数据防遗忘- LoRA 优先,效果好再考虑全参#AI #Agent #FunctionCalling #MCP #微调 #大模型
Related
Praca z najnowszym modelem Claude Fable 5 wymaga od programistów zmiany podejścia: zamiast precyzyjnych promptów, klucze...
Praca z najnowszym modelem Claude Fable 5 wymaga od programistów zmiany podejścia: zamiast precyzyjnych promptów, kluczem jest identyfikacja własnych martwych punktów i traktowanie...
#MissKitty knows that she is not the only expert and some days she's not even an expert.Who has some clues about #AI tha...
#MissKitty knows that she is not the only expert and some days she's not even an expert.Who has some clues about #AI that currently supports open #Research on #Politics?My experien...
=> the new browsing experience#ai #browsers #web #internet #technology #programming #business #companies #media #informa...
=> the new browsing experience#ai #browsers #web #internet #technology #programming #business #companies #media #information #datahttps://techcrunch.com/2026/07/03/as-the-browser-w...