README.md

---
pretty_name: Cloud-OpsBench
language:
- en
license: mit
tags:
- cloud-ops
- root-cause-analysis
- kubernetes
- agentic-evaluation
- microservices
size_categories:
- n<1K
configs:
- config_name: default
  default: true
  data_files:
  - split: benchmark
    path: index.jsonl
---

# Cloud-OpsBench — Official Dataset Distribution

This is the official dataset distribution of **Cloud-OpsBench**, maintained by
**LLM4Ops**. The canonical code and benchmark implementation repository is
[LLM4Ops/Cloud-OpsBench on GitHub](https://github.com/LLM4Ops/Cloud-OpsBench).
This Hugging Face repository distributes the benchmark artifacts, not the agent
or evaluation implementation.

Cloud-OpsBench evaluates agentic root cause analysis in Kubernetes-based cloud
systems. Each incident is a preserved state snapshot, allowing diagnostic tools
to replay recorded evidence without a running cluster. A correct joint diagnosis
identifies both the faulty component and its root cause.

## Release and provenance

- Distribution version: **v1.0.0**.
- Source GitHub commit: [`ea05dafbc87b5985df8721622d63d4229c2b4224`](https://github.com/LLM4Ops/Cloud-OpsBench/tree/ea05dafbc87b5985df8721622d63d4229c2b4224).
- Cases: **754** — Online Boutique **550**, TrainTicket **204**.
- `release.json` records source identity, per-system/category counts, and data-tree IDs.
- `manifest.json` records each distributed file's size and SHA256, excluding the manifest itself.

The three artifact directories are exported byte-for-byte from the specified
GitHub commit. The card, case index, and release manifests are distribution
metadata. Version numbers identify dataset distributions, not new experimental
splits or benchmark scores. Other redistributed or transformed copies should
identify this source and the version/commit they derive from.

## Contents

```text
benchmark/<system>/<category>/<case_id>/
    metadata.json
    tool_cache.json
    raw_data/
    code/                     # Online Boutique diagnostic source files
golden-trajectory/<system>/<category>/<case_id>/
    path1.json
    path2.json
process-label/<system>/<category>/<case_id>/
    milestone.json
```

- **Benchmark snapshots:** incident query, namespace, difficulty, ground-truth
  component/root cause, cached diagnostic-tool responses, Kubernetes states,
  logs, alerts, available metrics, and application source where present.
- **Process labels:** milestone annotations, admissible evidence, and dependency
  edges for process evaluation.
- **Golden trajectories:** two auxiliary expert diagnostic trajectories per
  incident. They are not the input to the current milestone-based process scoring.

`index.jsonl` has one row per incident, including `case_key`, `system`, `category`,
`case_id`, `namespace`, `query`, `difficulty`, `fault_taxonomy`, `fault_object`,
`root_cause`, and relative paths to the metadata, tool cache, milestone, and two
golden trajectories. The Dataset Viewer displays this index only, not a merged
interpretation of all the heterogeneous JSON artifacts.

## Download and use

Install `huggingface_hub`, then download the complete snapshot:

```python
from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="LLM4Ops/Cloud-OpsBench",
    repo_type="dataset",
    revision="main",  # For reproducible experiments, pin an HF commit or release tag.
    local_dir="cloudopsbench-data",
)
```

Use the tools, agents, MCP interface, and evaluator from the
[GitHub implementation](https://github.com/LLM4Ops/Cloud-OpsBench).
After cloning that repository, the three downloaded artifact directories can be
placed at its root, preserving their original relative paths. Match the dataset's
source commit to the implementation when reproducing a release.

For a lightweight index without downloading every snapshot:

```python
from datasets import load_dataset

cases = load_dataset("LLM4Ops/Cloud-OpsBench", split="benchmark")
```

Loading the index does **not** download the incident files referenced by each row;
use `snapshot_download` for the replay artifacts. All released cases are exposed
as a single `benchmark` split. No official training/validation/test division is
introduced by this distribution.

## Evaluation boundary and limitations

Task agents should receive the incident query/namespace and acquire observations
through diagnostic tools. Ground-truth `result` fields, milestone labels, and
golden trajectories are evaluator/research resources; do not expose them to an
agent during benchmark evaluation. The distribution is public, but a runner
must enforce this separation.

These are recorded incidents from two microservice systems, not a representative
sample of every production environment. Snapshot replay does not test live
remediation or changing cluster state. Use the GitHub implementation for the
diagnostic tool semantics and evaluation protocol.

## License

The project's MIT license is reproduced in `LICENSE`. Original notices in
embedded third-party source files are retained; distribution does not replace
their original licenses or claim authorship of that source.

## Citation

```bibtex
@article{wang2026cloud,
  title={Cloud-OpsBench: A reproducible benchmark for agentic root cause analysis in cloud systems},
  author={Wang, Yilun and Yu, Guangba and Huang, Haiyu and Huang, Yujie and Wang, Zirui and Chen, Pengfei and Lyu, Michael R},
  journal={arXiv preprint arXiv:2603.00468},
  year={2026}
}
@inproceedings{yu2026freezing,
  title={Freezing the Crime Scene: A State Snapshot Paradigm for Reproducible Agentic SRE Evaluation},
  author={Yu, Guangba and Wang, Yilun and Lyu, Michael R},
  booktitle={Proceedings of the 34th ACM International Conference on the Foundations of Software Engineering},
  pages={1733--1737},
  year={2026}
}
```

For reproducibility, also report the HF revision or distribution version and
source commit from `release.json`. Report issues and propose case changes in the
[canonical GitHub repository](https://github.com/LLM4Ops/Cloud-OpsBench/issues).