Weyaxi· @Weyaxi · X·· 1 天前AI 评分
TermGrade 开源发布：1k 个终端智能体 RL 环境与训练配方
Weyaxi 团队发布 TermGrade，这是一套面向终端智能体的高质量 RL 环境，完全开源，同时公开构建与评分的完整过程以及所用的 RL 训练配方。该数据集包含 1k 个可执行环境，每个都经执行验证并对照 6 个模型评分，附带支撑评分的全部 36k 条轨迹（含失败样本）。作者称用这些任务训练 gemma-4-31B-it，在其原本能解出一半的任务上使 Terminal-Bench 2.1 提升 +3.1。
We're releasing TermGrade at @aiand_: high-quality RL environments for terminal agents, fully open source, along with the full story of how we built and graded them, and the RL recipe we trained with.
That's 1k executable environments, each verified by execution and graded against 6 models, plus all 36k trajectories behind the grades, failures included.
And training gemma-4-31B-it on the tasks it solved half the time took it +3.1 on Terminal-Bench 2.1.
Thread below 🧵
来源：Weyaxi · x.com