README.md

metadata

license: cc-by-4.0
language:
  - en
tags:
  - hallucinations
  - synthetic-data
  - wikipedia
  - safety
  - nlp

Fake Wikipedia

A synthetic and semi-synthetic dataset designed for studying hallucinations, omissions, and factual inconsistencies in language models.

This dataset contains 10000 paragraphs derived from the agentlans/wikipedia-first-paragraph dataset, filtered for lengths between 1000 and 8000 characters.

Using Qwen/Qwen3.5-4B, two distinct text variants were generated for each entry based solely on the article title:

1. Fully Synthetic (fake): A completely hallucinated/made-up Wikipedia introduction generated entirely by the model.
2. Semi-Synthetic (edited): A stylistic rewrite of the original Wikipedia paragraph adapted to match the model's synthetic writing style.

📂 Data Fields

Column	Type	Description
id	string	Unique Wikipedia page identifier.
url	string	Direct link to the source Wikipedia article.
title	string	Title of the Wikipedia article.
text	string	The original, unaltered first paragraph from Wikipedia.
edited	string	The rewritten version of the original text matching the synthetic style.
fake	string	A completely fabricated Wikipedia-style paragraph generated by the model.

💡 Example Row

{
  "id": "12519725",
  "url": "https://en.wikipedia.org/wiki/Esp%C3%ADritu%20Santo%20antelope%20squirrel",
  "title": "Espíritu Santo antelope squirrel",
  "text": "The Espíritu Santo antelope squirrel (Ammospermophilus insularis) is a species of antelope squirrel in the family Sciuridae...",
  "edited": "The Espíritu Santo antelope squirrel (*Ammospermophilus insularis*) is a species of squirrel in the family Sciuridae...",
  "fake": "The Espíritu Santo antelope squirrel (*Sciurus granatensis*) is a species of rodent belonging to the genus *Sciurus*..."
}

⚠️ Limitations & Biases

• Content Divergence: The original (text) and fully synthetic (fake) paragraphs are not directly parallel in content or level of detail.
• Source Quality: Wikipedia lead paragraphs vary significantly in writing style, formatting, and structural quality.
• Topic Bias: The underlying source data inherits natural Wikipedia biases regarding which subjects have extensive, detailed lead paragraphs.
• Realistic Fakes: This dataset is for studying misinformation and hallucinations, not propagating them.

Licence

Creative Commons Attribution 4.0