AlexWortega/AgentTrove

AgentTrove-Soyuz数据集是open-thoughts/AgentTrove数据集的一个经过精心筛选的子集，专注于前沿智能体交互轨迹。该数据集收集了在terminus-2/Harness框架下生成的智能体交互轨迹，原始数据集包含170万行数据，经过过滤后保留了高质量部分。过滤过程包括：仅保留前沿教师模型（如GLM-4.7、GLM-4.6、Kimi K2.0 Thinking、MiniMax M2.0、GPT-OSS-120B），移除小型/中型模型；移除不相关的数据源（如nl2bash、bash textbook、gsm8k、puzzles）；移除因超时、错误等导致的崩溃轨迹；并应用了质量过滤器（如循环检测、错误尾部、重复轨迹、消息过短）。数据集以ChatML格式组织，每条轨迹包含一系列用户和助手之间的消息交换，助手消息通常包含JSON格式的分析和命令字段。数据集中还包含元数据，如教师模型、来源、结果等。该数据集适用于文本生成任务，特别是智能体行为的监督微调（SFT），并提供不同配置（如按教师模型划分的clean、unresolved和bad子集）以便于使用。数据集规模在10万到100万行之间，许可证为Apache-2.0。

The AgentTrove-Soyuz dataset is a curated subset of the open-thoughts/AgentTrove dataset, focusing on frontier agentic interaction traces. It contains agentic interaction traces generated in the terminus-2/Harness framework, with the original dataset comprising 1.7 million rows, filtered to retain high-quality portions. The filtering process includes: keeping only frontier teacher models (e.g., GLM-4.7, GLM-4.6, Kimi K2.0 Thinking, MiniMax M2.0, GPT-OSS-120B) while removing small/medium models; removing irrelevant data sources (e.g., nl2bash, bash textbook, gsm8k, puzzles); eliminating crashed traces due to timeouts, errors, etc.; and applying quality filters (e.g., loop detection, error tails, duplicate trajectories, too short messages). The dataset is organized in ChatML format, where each trajectory consists of a series of message exchanges between users and assistants, with assistant messages often containing JSON-formatted analysis and command fields. It also includes metadata such as teacher model, source, and result. This dataset is suitable for text-generation tasks, particularly supervised fine-tuning (SFT) for agentic behavior, and offers various configurations (e.g., clean, unresolved, and bad subsets divided by teacher model) for ease of use. The dataset size ranges between 100K and 1M rows, and it is licensed under Apache-2.0.