arena-env

Compact terminal tasks for training small coding agents, spread over eight domains (software engineering, industrial and physical systems, natural science, office work, finance and economics, mathematics and formal reasoning, cybersecurity, media production).

Each task gives the agent a few generated input files in /root/work, a precise specification of one output file, and a terminal (exec, write_file, evaluate actions over the Terminal-Bench 2 action schema). The reward is 1 when the task's pytest verifier passes against the produced output and 0 otherwise; there is no partial credit. Inputs are regenerated deterministically at every reset by gen_data.py.

tasks.jsonl lists the tasks (task_id, category, source); each task's folder holds task.md, gen_data.py, verifier/test_outputs.py and a reference oracle/solve.sh. The tasks were derived from the public SkillsBench tasks (benchflow/skillsbench) as smaller siblings exercising the same skills.