open-thoughts/OpenThoughts-Agent-SFT-100K
收藏资源简介:
OpenThoughts-Agent-SFT-100K是一个开源数据集,旨在为训练智能体提供高质量数据。它是OpenThoughts-Agent SFT扩展阶梯的一部分,包含100,000个示例,每个示例由任务描述和智能体轨迹对组成,用于微调OpenThinkerAgent-8B-SFT-100K和OpenThinkerAgent-32B-SFT-100K模型。该数据集是论文中描述的最终SFT数据集。任务来源包括SWE-Smith、StackExchange-SuperUser、StackExchange-Tezos(通过合成增强以扩展任务多样性)和IssueTasks。智能体轨迹由GLM-4.7-AWQ作为教师在terminus-2框架中生成,并经过过滤,只保留至少包含5个模型轮次的轨迹。数据集字段包括多轮对话轨迹、任务描述、轨迹来源、智能体元数据以及滚动记录信息。数据集规模为100,000行,使用GLM-4.7-AWQ作为教师模型和terminus-2作为框架。
OpenThoughts-Agent-SFT-100K is an open-source dataset curated for training agents. It is the 100,000-example point of the OpenThoughts-Agent SFT scaling ladder, containing (task, agent-trajectory) pairs used to fine-tune OpenThinkerAgent-8B-SFT-100K and OpenThinkerAgent-32B-SFT-100K models. This 100K set is the final OpenThoughts-Agent SFT dataset described in the paper. Tasks are drawn from the Top-4 sources: SWE-Smith, StackExchange-SuperUser, StackExchange-Tezos (synthetically augmented to expand task diversity), and IssueTasks. Agentic trajectories are generated by GLM-4.7-AWQ acting as the teacher in the terminus-2 harness, then filtered to traces with at least 5 model turns. The dataset includes fields such as conversations, task description, trace source, agent metadata, and rollout bookkeeping. It consists of 100,000 rows, with GLM-4.7-AWQ as the teacher model and terminus-2 as the harness.




