Datasets:

mohamedmady
/

Academic-Text-arxiv-gpt-gemini

license: mit
task_categories:
  - text-classification
language:
  - en
tags:
  - ai-generated-text-detection
  - machine-generated-text
  - academic-text
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files: academic_text.parquet

Academic Text: human vs AI-generated academic paragraphs

669,008 academic paragraphs for AI-generated text detection in the STEM / academic domain.

Composition

source	label	rows
`arxiv` (papers published before 2022)	0 (human)	469,008
`gpt-3.5-turbo`	1 (AI)	100,000
`gemini-2.0-flash`	1 (AI)	100,000

Human paragraphs are drawn from arXiv papers published before 2022, which predates widespread LLM authorship and so guarantees human origin. AI paragraphs were generated through the official APIs on STEM topics matched to the distribution of the human paragraphs.

Columns

text (string) -- the paragraph
label (int) -- 0 human, 1 AI-generated
source (string) -- arxiv, gpt-3.5-turbo, or gemini-2.0-flash

Usage

from datasets import load_dataset
ds = load_dataset("mohamedmady/Academic-Text-arxiv-gpt-gemini")["train"]
print(ds[0])

Notes

The class balance is roughly 70 % human / 30 % AI. Report balanced accuracy or class-wise recall rather than raw accuracy.

The original blank-line-separated .txt files are retained for provenance, but the parquet is the canonical form: the text files carry no labels.

Licence

The generated paragraphs are released under MIT. The human paragraphs derive from arXiv, whose papers carry their own per-paper licences; check the terms of the source papers before redistributing that portion.

Citation

@misc{academictext2026,
  title  = {Academic Text: A large-scale academic-domain benchmark for AI-generated text detection},
  author = {Mady, Mohamed and Reschke, Johannes and Schuller, Bj\"orn},
  year   = {2026},
  howpublished = {Hugging Face dataset}
}