Key Info
TermGrade released as high-quality, fully open-source RL environments for terminal agents, with 1,000 executable environments graded against 6 models and all 36k trajectories shared.
Highlights
- 1k executable environments, each verified by execution and graded against 6 models
- Full set of 36k trajectories published, including 14k failures
- Training gemma-4-31B-it on tasks it solved half the time yielded +3.1 on Terminal-Bench 2.1
- Environments, trajectories, exact training split, checkpoint, and blog all open on Hugging Face