README.md

---
language:
- nl
license: apache-2.0
task_categories:
- question-answering
- text-generation
size_categories:
- n<1K
tags:
- dutch
- government
- benchmark
- legal
- evaluation
- netherlands
pretty_name: DutchGovBench
---

# DutchGovBench v0.1

Evaluation benchmark for Dutch government AI systems. 100 questions across 9 categories, testing knowledge of Dutch law and public administration.

## What is this?

DutchGovBench tests whether AI models can accurately answer questions about Dutch government topics: social support law (Wmo 2015), youth law (Jeugdwet), participation law (Participatiewet), administrative law (Awb), municipal policy, objection procedures, privacy/GDPR, administrative oversight, and subsidies.

Each question includes expected legal references and a verified answer.

## Dataset structure

```json
{
  "id": "wmo_001",
  "category": "Wmo 2015",
  "question": "Wat is het doel van de Wmo 2015?",
  "expected_refs": ["art. 1.1.1 Wmo 2015"],
  "gold_answer": "De Wmo 2015 regelt...",
  "difficulty": "easy",
  "style": "factual",
  "verifiable": true
}
```

Fields:
- **id**: Unique question identifier
- **category**: One of 9 topic categories
- **question**: Question in Dutch
- **expected_refs**: Expected legal article references
- **gold_answer**: Verified reference answer
- **difficulty**: easy / medium / hard
- **style**: factual / applied / analytical
- **verifiable**: Whether answer can be checked against source text

## Baseline results

| Model | Avg Score (-2 to +2) | Hallucination Rate |
|-------|---------------------|-------------------|
| EuroLLM-9B-Instruct | -0.82 | 32% |
| EuroLLM-22B-Instruct | -0.93 | 55% |

## License

Apache 2.0

## Author

CiviQs B.V. (info@civiqs.nl)