CooperBench/qwen9b-solo-mini-swe-agent
收藏资源简介:
该数据集是通过在CooperData任务集上以solo(单智能体)模式运行CooperBench生成的单智能体编码轨迹。具体使用了Qwen/Qwen3.5-9B模型和mini_swe_agent_v2作为智能体框架。每个任务中,一个智能体负责实现两个功能。数据集包含368个任务,总令牌数(输入+输出)约为1.13亿。其中,两个功能均通过的任务比例为15.5%。数据集旨在与对应的coop(协作)版本数据集进行比较,以隔离合作缺陷。文件结构包括一个简化的索引文件(index.csv)和一个包含原始运行数据的目录(solo/),其中包含智能体会话事件、评估结果、轨迹和补丁等详细信息。数据集主要用于代码生成和智能体轨迹分析的研究。
This dataset contains single-agent coding trajectories generated by running CooperBench in solo mode on the CooperData task set, using the Qwen/Qwen3.5-9B model and the mini_swe_agent_v2 agent framework. In each task, one agent implements both features. The dataset includes 368 tasks with a total of approximately 113 million tokens (input + output). The pass rate for both features is 15.5%. It is designed to be compared with its matched coop version to isolate the cooperation deficit. The file structure consists of a slim index file (index.csv) and a raw run tree directory (solo/) containing detailed agent session events, evaluation results, trajectories, and patches. The dataset is intended for research on code generation and agentic trajectories.




