Long-horizon terminal tasks are hard to assemble at scale because correct solutions, verifiers, and runnable environments must remain consistent. This dataset addresses that bottleneck by publishing 37,484 validated command-line task instances produced via recursive task synthesis and executable sandbox validation, giving researchers a large corpus of verifiable, runnable terminal tasks for agent training and evaluation.
What Sets It Apart
- Verified runnable packages: each task is shipped with a sanitized runnable package (data shards of tasks-*.tar), a shard manifest, and searchable metadata rows (metadata/tasks.parquet) so tasks can be executed and validated in isolated sandboxes.
- Complete metadata per task: searchable fields include instruction (instruction.md), task_toml, solution (solution/solve.sh) when present, and environment/Dockerfile when present — enabling automated dataset filtering and trial reproduction.
- Recursive synthesis pipeline: tasks were generated by extending verified seed tasks through multiple synthesis rounds, producing increasingly long and complex reference solutions while maintaining end-to-end validation; this yields long-horizon workflows useful for training agents that must plan and execute many dependent shell commands.
- Open license and reproducibility: published under CC BY 4.0 and packaged with checksums and TAR manifests to support reproducible downloads and sandboxed evaluation.
Who It's For and Trade-offs
Great fit if you need a large collection of verifiable command-line benchmarks for training or evaluating terminal agents, reinforcement-learning policies, or executor components of agents — especially when you require runnable tasks and metadata for automatic validation, trajectory collection, or curriculum construction. Look elsewhere if you need non-command-line modalities (e.g., purely visual or conversational tasks), human-curated natural tasks at scale (these are synthetically expanded from seeds), or if executing tasks in a sandboxed environment is impractical for your setup.
Where It Fits
Use this dataset as a source of long-horizon, executable training data for RL or agent fine-tuning (supervised or RL-based), for benchmarking terminal-agent capabilities, or for research into task generation and verifier design. Combine with trajectory-collection workflows to produce agent rollouts; pair with sandbox orchestration to ensure safe execution and validation.
Notes: dataset layout emphasizes two top-level components — metadata (metadata/tasks.parquet and metadata/shard_manifest.jsonl) for discovery and integrity, and data shards (data/tasks-*.tar) containing sanitized runnable packages — enabling both scale and reproducibility without exposing unsafe or unvetted artifacts.
