Dressage-Claw

Dressage-Claw is a synthetic collection of 441 tool-use tasks for black-box agent reinforcement learning and evaluation. It is designed for Accio-Lab/Dressage, with OpenClaw as the agent harness and deterministic local mock HTTP services as the execution environment.

Each task bundles its prompt, tool definitions, service fixtures, lifecycle scripts, workspace, and grader. Tasks require agents to retrieve, reconcile, or update state while respecting explicit safety and authorization constraints.

Dataset overview

The default Hugging Face split is a 441-row task index. The complete executable task environments are stored in archive/Dressage-Claw.parquet.

Repository contents

Dressage-Claw/
├── README.md
├── LICENSE
├── THIRD_PARTY_NOTICES.md
├── manifest.json
├── data/
│   └── task_index.parquet
├── archive/
│   └── Dressage-Claw.parquet
└── scripts/
    └── unpack_archive.py

Task index

The Dataset Viewer and datasets.load_dataset() expose one row per task:

from datasets import load_dataset

index = load_dataset("huang3eng/Dressage-Claw", split="train")
print(index[0]["task_id"], index[0]["services"])

Index fields include the task ID, name, prompt, difficulty, services, tool count, primary grading dimensions, workspace and judge requirements, grader kind, maximum turns, execution budget, source labels, and archive path.

Download and unpack

Private-repository access requires prior authentication:

pip install -U huggingface_hub pyarrow
hf auth login
hf download huang3eng/Dressage-Claw \
  --repo-type dataset \
  --local-dir Dressage-Claw-HF

python Dressage-Claw-HF/scripts/unpack_archive.py \
  Dressage-Claw-HF/archive/Dressage-Claw.parquet \
  --output Dressage-Claw

To verify every path and SHA-256 digest without extracting:

python Dressage-Claw-HF/scripts/unpack_archive.py \
  Dressage-Claw-HF/archive/Dressage-Claw.parquet \
  --verify-only

The archive is a filesystem table with the following columns:

After extraction, each task is laid out as:

tasks/<task_id>/
├── task.json
├── task/
│   ├── prompt.md
│   ├── tools/
│   ├── services/
│   ├── scripts/
│   ├── runtime/
│   └── workspace/
└── grader/

Evaluation and Dressage integration

Every task contains fixed mock-service data and an independent grader. Graders primarily check whether required results are complete, whether tools were used correctly, and whether prohibited unsafe or unauthorized actions occurred.

An Omni Task-compatible runner is responsible for mounting the workspace, starting mock services, recording tool traces, and invoking the grader. Dressage-Claw is intended as an input bundle for Dressage black-box rollouts. The public Dressage repository does not currently ship a Dressage-Claw-specific loader, so training integration must supply the Omni Task runner and convert task records into the JSONL/configuration expected by the selected Dressage recipe.

Safety, privacy, and limitations

• Names, contact details, financial records, and credential-shaped values in service fixtures are synthetic mock data. They must not be treated as real identities, accounts, or secrets.
• Some fixtures deliberately contain sk-...- and xox...-shaped placeholder values to exercise configuration and safety behavior.
• The bundle vendors PyYAML 6.0.3 inside task environments, including a macOS Python 3.12 native extension. Linux runners should use the pure-Python fallback or rebuild dependencies for their platform.
• The collection has no official train/test separation; the single train split is an index over all 441 tasks.
• Results depend on the agent harness, model, sandbox, runner, and reward conversion used with Dressage.

Provenance

Dressage-Claw retains all 441 tasks from by_tier_omni/0621hard. See manifest.json for the complete task index, service statistics, and source metadata.

Citation

@misc{dressage_claw_2026,
  author       = {Dressage Contributors},
  title        = {{Dressage-Claw}: Synthetic Tool-Use Tasks for Agentic Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/huang3eng/Dressage-Claw}}
}

@misc{dressage_github,
  author       = {Liangmeng Huang and Qingchuan Li and Hongwei Xue and Shilin Yan and {Dressage Contributors}},
  title        = {{Dressage}: Scalable {RL} for Any Agent and Any Sandbox},
  year         = {2026},
  howpublished = {\url{https://github.com/Accio-Lab/Dressage}}
}

License

Dressage-Claw is released under the Apache License 2.0. Vendored third-party components retain their own licenses; see THIRD_PARTY_NOTICES.md.

---

中文概览

Dressage-Claw 是一个包含 441 个合成工具调用任务的数据集，面向 Dressage 的黑盒智能体强化学习与评测。 每个任务包含用户提示、工具定义、本地 Mock HTTP 服务、固定测试数据和独立评分器。

Hugging Face 默认展示的 data/task_index.parquet 是 441 行任务索引；完整可执行环境位于 archive/Dressage-Claw.parquet，可使用 scripts/unpack_archive.py 安全解包并校验散列。 任务中的姓名、联系方式、财务记录及密钥样式值均为合成 Mock 数据，不应视为真实身份或凭据。