kalomaze/wiki-judge-balanced

---
license: cc-by-4.0
language:
- en
tags:
- judge
- constraint-compliance
- synthetic
- reasoning-traces
size_categories:
- 10K<n<50K
---

# wiki-judge-balanced

Balanced constraint-compliance judge dataset. 18.9k examples with GPT-4.1-mini judge verdicts on wiki summary constraint tasks.

Each row contains a constraint, article, model response, and the original judge's XML verdict (`<consideration>` + `<judgement>`). YES/NO balanced per constraint type.

## Fields

```json
{
"constraint": "Summarize using exactly 5 words...",
"constraint_type": "exact_5_words",
"article": "# Article title\n...",
"completion": "<think>...</think>\nfull model output",
"think": "extracted reasoning",
"answer": "text after </think>",
"judge_response": "<consideration>...</consideration>\n<judgement>YES</judgement>",
"reward": 1.0
}
```

## Constraint Types

17 constraint types covering word counts, sentence counts, bullet points, formatting, and structural requirements.