open-thoughts/OpenThoughts-Agent-v1-SFT

OpenThinker-Agent-v1是一个开源项目，旨在为训练智能体（agents）提供最佳数据集。该项目的第一版包括了数据集、模型和研究代码库。OpenThinker-Agent-v1模型专为智能体任务设计，如Terminal-Bench 2.0和SWE-Bench。该模型基于Qwen/Qwen3-8B进行后训练，首先在OpenThoughts-Agent-v1-SFT数据集上进行监督微调（SFT），然后在OpenThoughts-Agent-v1-RL数据集上进行强化学习（RL）。OpenThoughts-Agent-v1-SFT数据集包含约15,200条轨迹，来源于两个数据源：nl2bash（简单合成生成的格式化shell命令任务）和InferredBugs（由微软收集的C#和Java中的bug集合）。OpenThoughts-Agent-v1-RL数据集包含约720个任务，来源于nl2bash验证数据集。项目还包括了训练超参数和框架版本的详细信息。

OpenThinker-Agent-v1 is an open-source effort to curate the best datasets for training agents. The first release includes datasets, models, and the research codebase. The OpenThinker-Agent-v1 model is trained for agentic tasks such as Terminal-Bench 2.0 and SWE-Bench. It is post-trained from Qwen/Qwen3-8B, SFT-ed on the OpenThoughts-Agent-v1-SFT dataset, and then RL-ed on the OpenThoughts-Agent-v1-RL dataset. The OpenThoughts-Agent-v1-SFT dataset contains approximately 15,200 traces from two data sources: nl2bash (simple synthetically generated tasks for formatting shell commands) and InferredBugs (a set of bugs in C# and Java collected by Microsoft). The OpenThoughts-Agent-v1-RL dataset contains ~720 tasks drawn from the nl2bash verified dataset. The project also details the training hyperparameters and framework versions used.