AOS-Trajectories
收藏资源简介:
AOS-Trajectories是一个专门用于代码代理工具调用轨迹的数据集,旨在为大型语言模型(LLMs)在Manim动画生成任务上进行微调提供数据支持。该数据集包含两种主要数据格式:原始代理轨迹和OpenAI风格的多轮工具调用轨迹。原始轨迹数据记录了从用户提示开始到最终生成可成功运行的Manim Python脚本的完整工具调用序列,每个样本包括用户提示、工具步骤列表(含类型、工具名称、输入、输出和错误标志)、最终代码、可选的叙述摘要以及成功状态等字段。工具轨迹数据则被处理成适用于工具使用或代码模式微调的多轮对话格式,包含用户、助理和工具角色之间的交互消息。数据集规模小于1000个样本,主要用于监督微调(SFT)阶段,以提升LLMs在理解和执行Manim动画编程任务中的工具调用与代码生成能力。数据通过AOS代码代理收集并导出,适用于动画生成、代码生成和工具使用等研究与应用场景。
AOS-Trajectories is a dataset specifically designed for code agent tool-call trajectories, aimed at providing data support for fine-tuning large language models (LLMs) on Manim animation generation tasks. As part of the AOS project, it includes two main data formats: raw agent trajectories and OpenAI-style multi-turn tool-call trajectories. The raw trajectory data records the complete sequence of tool calls from user prompts to the final generation of a successfully executable Manim Python script, with each sample containing fields such as user prompts, a list of tool steps (including type, tool name, input, output, and error flags), final code, optional narrative summaries, and success status. The tool trajectory data is processed into a multi-turn dialogue format suitable for tool usage or code pattern fine-tuning, involving interaction messages among user, assistant, and tool roles. The dataset contains less than 1000 samples and is primarily used for the supervised fine-tuning (SFT) phase to enhance LLMs capabilities in tool calling and code generation for understanding and executing Manim animation programming tasks. Data is collected and exported via the AOS code agent, making it applicable to research and application scenarios such as animation generation, code generation, and tool usage.





