Datasets:

saulpzh
/

Meta-Task-Data

license: apache-2.0
language:
  - en
task_categories:
  - text-generation
tags:
  - terminal-agents
  - code-agent
  - task-synthesis
  - agent-trajectories
  - sft

Meta-Task-Data

Data produced by Meta-Task, the official implementation of "Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Training" (arXiv:2607.27929).

Meta-Task redefines terminal task synthesis as a Terminal-Bench-format terminal task: an agent runs inside a Docker container to generate, execute, and verify new terminal tasks.

Path	Contents
`tasks/meta_tasks_v1.tar.gz`	3,220 synthesized terminal tasks. Each is a complete Terminal-Bench package (`instruction.md`, `task.toml`, `environment/`, `solution/`, `tests/`). Extracts to `v1/<task-name>/`.
`trajectories/sft_trajectories_3221.jsonl`	3,221 SFT trajectories, one JSON object per line.

Trajectory format

Each line of sft_trajectories_3221.jsonl:

Field	Description
`messages`	Conversation turns (`role`, `content`) of the agent solving the task.
`id`	Name of the source task (matches a directory under `v1/`).
`data_type`	`sft`
`special_ability`	`code_agent`
`source`	`terminalbench`
`synthetic_model`	Model that generated the trajectory (Qwen3.5-397B-A17B-FP8).
`tag`	e.g. `{"code_agent": "terminus-2"}` — the agent scaffold used.

Loading

import json, tarfile

with tarfile.open("tasks/meta_tasks_v1.tar.gz") as tf:
    tf.extractall("tasks/")          # -> tasks/v1/<task-name>/

with open("trajectories/sft_trajectories_3221.jsonl") as f:
    trajectories = [json.loads(line) for line in f]

License

Apache-2.0.

Meta-Task-Data

Contents

Trajectory format

Loading

Links

License