README.md

---
dataset_info:
  features:
  - name: json_schema
    dtype: string
  - name: unique_id
    dtype: string
  splits:
  - name: WashingtonPost
    num_bytes: 2710348
    num_examples: 125
  - name: Snowplow
    num_bytes: 1613804
    num_examples: 403
  - name: Kubernetes
    num_bytes: 25623424
    num_examples: 1064
  - name: Github_trivial
    num_bytes: 780060
    num_examples: 444
  - name: Github_easy
    num_bytes: 1980784
    num_examples: 1943
  - name: Github_medium
    num_bytes: 7994298
    num_examples: 1976
  - name: Github_hard
    num_bytes: 20240875
    num_examples: 1240
  - name: Github_ultra
    num_bytes: 12235981
    num_examples: 164
  - name: JsonSchemaStore
    num_bytes: 22195651
    num_examples: 492
  - name: Glaiveai2K
    num_bytes: 1440707
    num_examples: 1707
  download_size: 19019152
  dataset_size: 96815932
configs:
- config_name: default
  data_files:
  - split: WashingtonPost
    path: data/WashingtonPost-*
  - split: Snowplow
    path: data/Snowplow-*
  - split: Kubernetes
    path: data/Kubernetes-*
  - split: Github_trivial
    path: data/Github_trivial-*
  - split: Github_easy
    path: data/Github_easy-*
  - split: Github_medium
    path: data/Github_medium-*
  - split: Github_hard
    path: data/Github_hard-*
  - split: Github_ultra
    path: data/Github_ultra-*
  - split: JsonSchemaStore
    path: data/JsonSchemaStore-*
  - split: Glaiveai2K
    path: data/Glaiveai2K-*
pretty_name: J
---

# JSONSchemaBench

[![Paper](https://img.shields.io/badge/Paper-arXiv-blue)](https://arxiv.org/abs/2501.10868)  

JSONSchemaBench is a benchmark of **real-world JSON schemas** designed to evaluate **structured output generation** for Large Language Models (LLMs). It contains approximately **10,000 JSON schemas**, capturing diverse constraints and complexities.

## 📌 Dataset Overview
- **Purpose:** Evaluate the **efficiency** and **coverage** of structured output generation.
- **Sources:** GitHub, Kubernetes, API specifications, curated collections.
- **Schemas:** Categorized based on complexity and domain.

### 📊 Dataset Breakdown
| Dataset         | Category            | Count |
| --------------- | ------------------- | ----- |
| GlaiveAI-2K     | Function Call       | 1707  |
| Github-Trivial  | Misc                | 444   |
| Github-Easy     | Misc                | 1943  |
| Snowplow        | Operational API     | 403   |
| Github-Medium   | Misc                | 1976  |
| Kubernetes      | Kubernetes API      | 1064  |
| Washington Post | Resource Access API | 125   |
| Github-Hard     | Misc                | 1240  |
| JSONSchemaStore | Misc                | 492   |
| Github-Ultra    | Misc                | 164   |
| **Total**       |                     | 9558  |

## 📥 Loading the Dataset

```python
from datasets import load_dataset

dataset = load_dataset("epfl-dlab/JSONSchemaBench")
print(dataset)
```

## 🔍 Data Structure
Each dataset split contains:
- `"json_schema"`: The schema definition.
- `"unique_id"`: A unique identifier for the schema.


🚀 **For more details, check out the [paper](https://arxiv.org/abs/2501.10868).**

## 📚 Citation
If you use this dataset, please cite:
```bibtex
@misc{geng2025jsonschemabench,
      title={Generating Structured Outputs from Language Models: Benchmark and Studies},
      author={Saibo Geng et al.},
      year={2025},
      eprint={2501.10868},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.10868}
}
```


## License

This dataset is provided under the [MIT License](https://opensource.org/licenses/MIT). Please ensure that you comply with the license terms when using or distributing this dataset.

## Acknowledgements

We would like to thank the contributors and maintainers of the JSON schema projects and the open-source community for their invaluable work and support.