open-thoughts/OpenThoughts-Agent-SFT-ColdStartForRL-10K
收藏资源简介:
OpenThoughts-Agent-SFT-ColdStartForRL-10K是一个开源数据集,属于OpenThoughts-Agent项目的一部分,旨在为训练智能体(agents)提供最佳数据集。该数据集包含9,437个(任务,智能体轨迹)对,用于在强化学习(RL)之前进行冷启动监督微调(SFT),以教导基础模型智能体的交互格式和工具使用行为,从而使后续的强化学习更加稳定和样本高效。数据来源于SWE-Smith的沙盒编码任务,这些任务带有测试,并在terminus-2框架和Daytona沙盒中生成轨迹,经过验证(验证超时120秒)。每个数据行存储为多轮对话轨迹(conversations字段),采用角色/内容的消息格式(系统/用户/助手),适合直接用于监督微调。该数据集用于微调Qwen/Qwen3-8B模型,生成OpenThinkerAgent-8B-ColdStartSFTForRL模型,作为强化学习前的基座。
OpenThoughts-Agent-SFT-ColdStartForRL-10K is an open-source dataset curated as part of the OpenThoughts-Agent effort to provide the best datasets for training agents. It contains 9,437 (task, agent-trajectory) pairs designed for cold-start supervised finetuning (SFT) to teach a base model the agentic format and behavior before reinforcement learning (RL), making subsequent RL stable and sample-efficient. The data is sourced from SWE-Smith sandboxed coding tasks with tests, generated in the terminus-2 harness inside Daytona sandboxes, and oracle-verified with a 120s verifier timeout. Each row is stored as a multi-turn agent trajectory in the `conversations` chat format with role/content messages (system/user/assistant), suitable for direct supervised finetuning. The dataset is used to fine-tune Qwen/Qwen3-8B to produce OpenThinkerAgent-8B-ColdStartSFTForRL, which serves as the pre-RL base.




