txchmechanicus/SWE-ZERO-12M-trajectories

SWE-ZERO-12M-trajectories是一个大规模代理编码轨迹数据集，包含1120亿令牌的无执行代理轨迹，覆盖122,908个独特的拉取请求、3,222个代码仓库和16种编程语言。该数据集旨在通过规避Docker容器化执行的限制，提高数据可扩展性和训练效率，使用无执行管道从真实GitHub PR快照中采样，通过模型生成多轮轨迹，遵循mini-swe-agent v1格式。数据包括实例ID、仓库信息、消息轨迹、轨迹格式、退出状态和生成时间等字段。采样设置包括每轮最大15个回合、每个PR 100个轨迹、温度1.0等。数据集主要用于中训练，以培养模型的代理工具使用先验，但未经验证，大多数轨迹未成功提交，且受15回合限制。实验验证显示其能提升模型在SWE-bench等任务上的性能。

SWE-ZERO-12M-trajectories is a large-scale agentic-coding trace dataset containing 112 billion tokens of execution-free agentic trajectories, covering 122,908 unique pull requests, 3,222 repositories, and 16 programming languages. It addresses scalability bottlenecks of containerized execution by using an execution-free pipeline that samples from real GitHub PR snapshots and generates multi-turn rollouts via a model, following the mini-swe-agent v1 format. The dataset includes fields such as instance_id, repo, messages, trajectory_format, exit_status, and duration_sec. Sampling settings involve a maximum of 15 turns per rollout, 100 rollouts per PR, temperature 1.0, and other parameters. Intended for mid-training, it aims to instill agentic tool-use priors in base models but lacks verification, with most rollouts not reaching successful submission and being truncated at 15 turns. Experimental validation shows it improves model performance on tasks like SWE-bench.