AlienKevin/SWE-ZERO-12M-trajectories

SWE-ZERO 12M轨迹数据集是一个包含无执行代理代码编辑轨迹的数据集，源自SWE-ZERO管道。数据集当前检查点为30B，包含2,795,433次滚动，27,749个唯一PRs，估计31.9B tokens，提交率为5.8%。使用的模型是ricdomolm/mini-coder-1.7b，数据源为nebius/SWE-rebench-V2-PRs（126K PRs，20种语言）。目标是12.3M滚动/140B tokens。数据集管道包括使用Qwen3-1.7B模型在400K mini-swe-agent轨迹上进行微调，基础设施为TPU v6e-4/v5p-8/v5litepod-4，配方为vLLM serve with TP=4，前缀缓存，32K上下文，并发=64，格式为mini-swe-agent v1（仅bash交互，沙盒执行），质量控制在生成时过滤错误滚动，并通过(instance_id, message_hash)去重。数据集每10B token里程碑更新一次。

The SWE-ZERO 12M Trajectories dataset contains execution-free agentic code-editing trajectories from the SWE-ZERO pipeline. The current checkpoint is 30B, with 2,795,433 rollouts, 27,749 unique PRs, an estimated 31.9B tokens, and a submission rate of 5.8%. The model used is ricdomolm/mini-coder-1.7b, and the dataset source is nebius/SWE-rebench-V2-PRs (126K PRs, 20 languages). The target is 12.3M rollouts / 140B tokens. The pipeline includes a Qwen3-1.7B model fine-tuned on 400K mini-swe-agent trajectories, infrastructure using TPU v6e-4/v5p-8/v5litepod-4, a recipe of vLLM serve with TP=4, prefix caching, 32K context, concurrency=64, format of mini-swe-agent v1 (bash-only interaction, sandboxed execution), and quality control with error rollouts filtered at generation time and deduplicated by (instance_id, message_hash). The dataset is updated at every 10B token milestone.