five

TeichAI/DeepSeek-v4-Pro-Agent

收藏
Hugging Face2026-05-22 更新2026-06-14 收录
下载链接:
https://hf-mirror.com/datasets/TeichAI/DeepSeek-v4-Pro-Agent
下载链接
链接失效反馈
官方服务:
资源简介:
DeepSeek v4 Pro Agent Traces数据集是一个由TeichAI使用teich工具生成的代理跟踪文件集合,专门用于监督微调训练。该数据集包含4006个JSONL文件,每个文件代表一个捕获的代理会话,以新行分隔的JSON格式存储。所有助手响应均由deepseek/deepseek-v4-pro模型生成,确保高质量和一致性。数据集设计为上传优先保存,允许保留原始会话历史,并可后续转换为训练格式。常见的事件组包括session_meta、turn_context、event_msg、response_item、session、message、session_info、model_change和thinking_level_change。该数据集适用于训练AI代理,支持工具调用(如bash、read_file、write_file),并提供了完整的工具模式快照,便于集成到训练流程中。使用示例代码,用户可以轻松加载和转换数据,结合Unsloth和TRL SFTTrainer进行高效训练。

The DeepSeek v4 Pro Agent Traces Dataset is a collection of agent trace files generated by TeichAI using the teich tool, specifically tailored for supervised fine-tuning training. This dataset contains 4006 JSONL files, each representing a captured agent session and stored in newline-delimited JSON format. All assistant responses are generated by the deepseek/deepseek-v4-pro model to guarantee high quality and consistency. The dataset is designed with upload-first preservation in mind, enabling retention of original session histories and subsequent conversion into training-ready formats. Common event groups include session_meta, turn_context, event_msg, response_item, session, message, session_info, model_change, and thinking_level_change. This dataset is applicable to AI Agent training, supports tool calls such as bash, read_file, and write_file, and provides complete tool schema snapshots to facilitate integration into training workflows. Users can easily load and transform the data using the provided sample code, and carry out efficient training in conjunction with Unsloth and TRL SFTTrainer.
提供机构:
TeichAI
二维码
社区交流群
二维码
科研交流群
商业服务