build-small-hackathon/AI-Puppet-Theater-Actor-SFT
收藏资源简介:
AI Puppet Theater Actor SFT是一个合成监督微调数据集,专为AI Puppet Theater / The Runaway Puppet Show中的Actor代理设计。该数据集用于教导小型语言模型以紧凑的JSON对象响应单个木偶剧场节拍,适用于黑客马拉松原型设计、模式遵循和本地适配器实验,不适用于通用讲故事或聊天任务。数据集中每个条目为聊天风格的JSONL格式,包含系统、用户和助理角色消息,其中助理响应为序列化的JSON对象,包括意图、台词、情感、手势、舞台效果、记忆更新和工具请求等字段。数据集通过确定性模板生成,基于固定种子、多个前提、演员配置等元素,并支持可选外部种子源。其用途包括微调小型模型以生成Actor代理JSON、测试结构化输出和工具请求形状、以及黑客马拉松原型开发。局限性包括合成和模板驱动、记忆弧线浅、仅限Actor代理等。
AI Puppet Theater Actor SFT is a synthetic supervised fine-tuning dataset specifically designed for the Actor agents in AI Puppet Theater / The Runaway Puppet Show. This dataset is intended to teach small language models to respond to individual puppet theater beats with compact JSON objects, and is tailored for hackathon prototyping, pattern adherence, and local adapter experiments, but not for general storytelling or chat tasks. Each entry in the dataset follows a chat-style JSONL format, containing system, user, and assistant role messages, where the assistant's response is a serialized JSON object including fields such as intent, lines, emotion, gestures, stage effects, memory updates, and tool requests. The dataset is generated using deterministic templates based on elements including fixed seeds, multiple premises, and actor configurations, and supports optional external seed sources. Its use cases include fine-tuning small models to generate Actor agent JSON, testing structured output and tool request formats, and hackathon prototype development. Its limitations include being synthetic and template-driven, having shallow memory arcs, and being exclusively limited to Actor agents, among others.




