kadenwren/terminal-bench

Solve end-to-end shell tasks inside an isolated container.

A verifiable task: an instruction, a container environment, and a hidden test script. Each trial runs in the sandbox, the verifier scores the final state and writes the task's reward file, and the trajectory published here is exactly what was scored.

Format

Harbor compatible

Reward

reward.txt or reward.json

Verifier

tests/, hidden

Trials

1 attempt each