hamishivi/swerl-combined-agent-tasks

---
license: apache-2.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: messages
list:
- name: content
dtype: string
- name: role
dtype: string
- name: ground_truth
dtype: string
- name: dataset
dtype: string
- name: env_config
struct:
- name: env_name
dtype: string
- name: image
dtype: string
- name: task_id
dtype: string
- name: source
dtype: string
splits:
- name: train
num_bytes: 10847711
num_examples: 4309
download_size: 4447921
dataset_size: 10847711
---

# SWERL Combined Agent Tasks

Combined dataset for SWERL sandbox RL training. Includes task data tarball with per-task Docker images.

## Sources

| Source | Rows | Description |
|--------|------|-------------|
| [hamishivi/agent-task-combined](https://huggingface.co/datasets/hamishivi/agent-task-combined) | 3,627 | Terminal automation tasks (endless_terminals, openthoughts, swe_gym) |
| [osieosie/tmax-tasks-skill-taxonomy-20260324-1k-verified](https://huggingface.co/datasets/osieosie/tmax-tasks-skill-taxonomy-20260324-1k-verified) | 682 | Skill taxonomy tasks with verified test suites |

**Total: 4,309 tasks**

## Format

Each row has:
- `messages`: system + user prompt
- `ground_truth`: task ID
- `dataset`: "passthrough"
- `env_config`: `{env_name, task_id, image}` — Docker image tag for the task
- `source`: origin dataset identifier

## Task Data

The `task-data.tar.gz` file contains per-task directories:
- `{task_id}/instruction.md` — task description
- `{task_id}/tests/test.sh` — test script (uploaded only at submit time)
- `{task_id}/setup.sh` — container setup commands
- `{task_id}/environment/seeds/` — seed files for /workspace