Short-horizon agent task sets: every task from the source benchmarks whose agent time cap ([agent] timeout_sec in task.toml) is at most 20 minutes (1200 s).
Each subset is a top-level directory of standard Harbor task dirs (task.toml, instruction.md, environment/, tests/), runnable with harbor run -p .
tblite: openthoughts/openthoughts-tblite (100 tasks).
terminal-bench-2: terminal-bench/terminal-bench-2 (89 tasks).… See the full description on the dataset page:
https://huggingface.co/datasets/yoonholee/agent-tasks-short.
