[GRADTEX - GRAded Detection of TEXts]

A benchmark for fine-grained AI-generated text detection. Unlike binary human-vs-machine datasets, this dataset represents the continuum of machine involvement in text production: from purely human-authored texts, through texts edited or partially rewritten by language models, to texts fully generated from scratch.

Dataset summary

The dataset is built on top of MAGE (Li et al., 2024) and extends it in two directions:

1. Filtered human-authored texts from MAGE are used as the source material for all editing and partial-generation scenarios.
2. Thirteen generation scenarios are applied to those texts using seven modern language models, producing variants with different degrees of machine involvement (token-level edits, sentence-level edits, paraphrases, completions, style rewrites, full topic-based generation). Nine of the thirteen scenarios are used in the train and validation splits; the remaining four appear only in the test split (concentrated in the Test C subset).

For Test A (pure HWT vs pure MGT) the dataset also includes 3 332 fully machine-generated texts directly from MAGE, restricted to the topical_prompt and specified_prompt generation setups (other MAGE generation setups are not included). These rows are marked in the dataset by:

• original_name_in_MAGE — identifier of the source row in MAGE;
• MAGE_scenario_name — the original MAGE prompt setup;
• evaluation_setup — the evaluation regime they belong to (test_A).

Source and license

The human-authored texts and a subset of machine-authored texts in this dataset are sourced from MAGE (Li et al., 2024), released under the Apache License 2.0. This derivative dataset extends MAGE with additional machine-generated and partially-edited texts produced by multiple language models, and is released under CC BY 4.0.

Data composition

Human-authored texts (from MAGE)

Human texts are inherited from MAGE and additionally filtered:

• Domain hswag excluded — removed entirely from the source.
• Language filtering — only English texts retained.
• Length filtering — texts that are too short or too long (statistical outliers in word count and sentence count) removed.
• Outlier removal — additional filtering of malformed and atypical samples.

Machine-generated texts from MAGE

Generation scenarios

The dataset contains 13 generation scenarios in total, grouped into five families. Train and validation use 9 of them; the remaining 4 are present only in the test split (concentrated in Test C) and serve as out-of-distribution transformation types unseen during training.

#	Scenario	Family	Used in train / val
1	Topic-based full generation	MGT	yes
2	Polish — token level	Polish	yes
3	Polish — sentence level	Polish	no (test only)
4	Paraphrase	Paraphrase	yes
5	Complete ending	Completion	yes
6	Complete beginning	Completion	yes
7	Complete middle	Completion	no (test only)
8	Rewrite style — informal	Style rewriting	yes
9	Rewrite style — emotional	Style rewriting	yes
10	Rewrite style — figurative	Style rewriting	yes
11	Rewrite style — skeptical	Style rewriting	no (test only)
12	Rewrite style — persuasive	Style rewriting	yes
13	Rewrite style — maximal	Style rewriting	no (test only)

Topic-based full generation

Topic generation. A two-step pipeline. First, key words and phrases are extracted from the human source text. Then a new text of comparable length and the same domain is generated from those keywords. The resulting text is fully machine-written, but thematically aligned with the human source.

Polish scenarios (minimal local edits)

Polish — token level. The model corrects spelling errors, minor grammatical issues, local word repetition, and out-of-context word choices via single-word substitutions and minimal grammatical adjustments. Sentence order and overall structure are preserved.

Polish — sentence level (test only). The model rewrites approximately one third of the sentences in a noticeable but localized way, preferring lexical or syntactic variation over outright correction. The fraction ⅓ was chosen because preliminary edit-distance measurements (cosine similarity, Levenshtein distance, Jaccard distance) showed almost no difference in transformation depth between 10 %, 20 %, and 30 % sentence-level rewriting rates, so the largest of the tested fractions was adopted.

Paraphrase (substantial rewriting)

Paraphrase. A substantial portion of the text is rewritten while preserving its original meaning. The model may reorder words within sentences, replace lexical items with appropriate alternatives, and modify syntactic structure, but the overall text structure remains recognizable.

Completion scenarios (regenerate one third of the text)

In all three completion scenarios, two thirds of the human text serve as context and the remaining one third is regenerated by the model. The context and the generated fragment are then concatenated to form the final sample.

Complete ending. Given the first ⅔ of the text, the model generates the final ⅓.

Complete beginning. Given the last ⅔ of the text, the model generates the missing opening ⅓ such that it leads smoothly into the provided continuation.

Complete middle (test only). Given the beginning and the ending (together ⅔ of the text), the model generates the missing middle ⅓ so that the full text is coherent.

Style rewriting scenarios

Rewrite style — informal. The text is rewritten in a casual, conversational register: contractions, simple everyday language, direct address to the reader, with optional slang, colloquialisms, internet memes, or emojis.

Rewrite style — emotional. The text is rewritten with expressive vocabulary, emotional coloring, vivid adjectives and adverbs, and varied sentence rhythm intended to evoke feelings while maintaining clarity.

Rewrite style — figurative. The text is rewritten with metaphors, similes, imagery, sensory details, and creative comparisons in place of literal descriptions.

Rewrite style — skeptical (test only). The text is rewritten with a questioning, analytical tone: cautious language, balanced evaluation of pros and cons, logical reasoning, and emphasis on evidence over assumptions.

Rewrite style — persuasive. The text is rewritten with rhetorical devices, benefit-oriented language, calls to action, emphasis on advantages, and a confident, motivational tone.

Rewrite style — maximal (test only). All of the preceding stylistic transformations are applied simultaneously — informal, emotional, figurative, persuasive, and skeptical — producing a maximally stylistically modified version of the text while preserving its original meaning.

Generator models

Seven language models were used across the dataset:

Model	Family
Qwen 3.5 27B	Qwen
Gemma 4 31B Instruct	Gemma
Gemma 4 E4B Instruct	Gemma
Mistral Small 3.2 24B Instruct 2506	Mistral
Claude Sonnet 4.6	Anthropic
GPT 5.4 Mini	OpenAI
Gemini 3.5 Flash	Google

Citation

If you use this dataset, please cite both the original MAGE paper and this work:

@misc{li2024mage,
      title={MAGE: Machine-generated Text Detection in the Wild}, 
      author={Yafu Li and Qintong Li and Leyang Cui and Wei Bi and 
              Zhilin Wang and Longyue Wang and Linyi Yang and Shuming Shi 
              and Yue Zhang},
      year={2024},
      eprint={2305.13242},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{your_dataset_2026,
      title={[Название твоего датасета]: [Подзаголовок]},
      author={Surname, Name},
      year={2026},
      howpublished={\url{https://huggingface.co/datasets/your-username/dataset-name}}
}