Datasets:

ErenJaegerYeager
/

ClawBenchPro

license: mit
language:
  - zh
  - en
task_categories:
  - text-generation
  - question-answering
pretty_name: ClawBenchPro
tags:
  - agent-benchmark
  - workplace
  - tool-use
  - multi-turn
  - skills
  - nanoclaw
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    default: true
    data_files:
      - split: test
        path: dataset_index.csv
  - config_name: round_01_aligned_mix_800
    data_files:
      - split: test
        path: round_01_aligned_mix_800/dataset_index.csv
  - config_name: persona_aligned_mix_200
    data_files:
      - split: test
        path: persona_aligned_mix_200/dataset_index.csv

ClawBenchPro

ClawBenchPro is a compact, builder-based workplace-agent benchmark package exported from Nanoclaw. It contains task YAML files, prompts, task-local environment builders, skills, evaluation manifests, provenance metadata, and checksums.

Included Splits

Dataset	Tasks	Groups
`round_01_aligned_mix_800`	800	`base`, `hard_aligned`, `multi_turn_aligned`, `skills_aligned`
`persona_aligned_mix_200`	200	`base`, `hard`, `multi_turn`, `skills`

Directory Layout

ClawBenchPro/
├── README.md
├── LICENSE
├── dataset_index.jsonl
├── manifest.json
├── checksums.sha256
├── round_01_aligned_mix_800/
└── persona_aligned_mix_200/

Each dataset directory contains:

tasks/: task YAML files, prompts, and task-local env_builder.py builders.
skills/: packaged skills referenced by task YAML files.
eval_manifests/: group-level manifests and task id lists.
provenance/: sanitized construction metadata.
manifest.json: dataset-level metadata.
checksums.sha256: dataset-level file checksums.

Prebuilt assets/ directories are intentionally not included to keep the Hugging Face repository compact. Each task includes an env_builder.py that can materialize assets/<task_id>/ on demand.

Usage

After downloading the dataset, point Nanoclaw or compatible runners at the YAML tasks under:

round_01_aligned_mix_800/tasks/*.yaml
persona_aligned_mix_200/tasks/*.yaml

Group manifests are available under eval_manifests/ for category-level analysis.

To materialize one task environment manually:

cd round_01_aligned_mix_800
python tasks/data_round_01_aligned_mix_800_0001/env_builder.py

This creates:

round_01_aligned_mix_800/assets/data_round_01_aligned_mix_800_0001/

Nanoclaw's batch runner can also invoke these builders automatically before each task run.

To materialize assets in batches from the repository root:

python materialize_assets.py --dataset round_01_aligned_mix_800 --workers 8
python materialize_assets.py --dataset persona_aligned_mix_200 --workers 8

Dataset Index

dataset_index.jsonl provides one row per task with:

dataset, task_id, category, task_file, asset_dir, prompt_files, skill_count

The full task definitions remain in the YAML files.

License

This package is released under the MIT License. See LICENSE.

Notes

The package is intended as a benchmark artifact rather than a tabular training dataset.
Some tasks intentionally contain synthetic keys, internal URLs, noisy logs, broken files, or policy-sensitive strings as part of the benchmark environment. These are benchmark fixtures, not operational credentials.
Build-time local absolute paths have been removed from the Hugging Face-ready package.