Ru-hard-detection-dataset

A dataset for evaluating the detection of generated texts in Russian language. It collects original human texts, AI generated versions and paraphrased variants across several genres.

Dataset Structure

The root directory contains two main directories:

• main – the core part of the dataset, containing three themes:

  • news – news articles
  • essay – essays
  • scientific_texts – scientific articles

• long_sc – additional scientific research texts

Each JSON file is a list of records with the following fields:

• id – unique integer identifier
• text – text of the example
• dataset – origin of the text (e.g., Lenta.ru news)
• source – human for original texts, ai for generated ones and ai+rew for paraphrased variants
• model – generating model name (only for generated and paraphrased files)
• paraphrasing_type – paraphrase strength: LP (light) or DP (deep) where present

Each of these folders contains three files:

1. Original
2. Generated text – generated based on the summarization of the original. In other words, key information was extracted from the original text and used as a basis for generating new, similar text.
3. Paraphrased text – rewritten original using two variants: weak and strong paraphrasing.

Location	File	Records
main/news	original_news.json	480
	generated_news.json	480
	paraphrased_news.json	480
main/essay	original_essay.json	480
	generated_essays.json	480
	paraphrased_essays.json	480
main/scientific_texts	orig_scientific.json	479
	generated_scientific.json	479
	paraphrased_scientific.json	479
long_sc	rus_scientific_articles_part1.json	1239
	rus_scientific_articles_part2.json	1239
	generated_articles.json	1449
	paraphrased_generated_articles.json	1625

Models and Sources

The JSON entries keep track of the origin of each text and the model that produced it. Generation was performed from short summaries of the originals using the following models: gemini-2.0-flash, chatgpt-4o-mini, chatgpt-o1-mini, deepseek-chat and deepseek-reasoner. Paraphrases were created with the same models in two modes – LP (light) and DP (deep). Original texts were collected from Lenta.ru news articles, the Corus Essays corpus, a Russian summarization dataset and a collection of Russian scientific articles.

Example Usage

The script examples/load_dataset_example.py iterates over all JSON files in the main directory and prints one full record from each. Run it with Python:

python examples/load_dataset_example.py