---
license: mit
task_categories:
- text-classification
language:
- en
tags:
- AI-Generated
- Human-Generated
- Academic
pretty_name: AI-Generated Academic Text Detection
size_categories:
- 100K<n<1M
---
language: en
license: cc-by-4.0
task_categories:
  - text-classification
tags:
  - ai-generated-text-detection
  - academic-text
  - distribution-shift
---

# Academic Text Dataset

**A large-scale academic-domain benchmark for AI-generated text detection.**

This dataset was specifically constructed to evaluate robustness of AI-text detectors under realistic distribution shift in the academic (STEM) domain.

## Dataset Description

The dataset contains:
- **469,008** human-written arXiv abstracts (collected from papers published **before 2022**, guaranteeing purely human-generated text)
- **100,000** AI-generated abstracts produced by **GPT-3.5 Turbo**
- **100,000** AI-generated abstracts produced by **Gemini 2.0 Flash**

**Total: 669,008 samples**

The AI texts were generated using the official APIs on STEM topics carefully matched to the distribution of the human arXiv abstracts, ensuring close semantic and topical alignment.

## Usage in Our Paper

We evaluate our best model (`DeBERTa-v3-base+FeatAttn`) under a **fixed-threshold protocol** (threshold calibrated once on HC3 PLUS validation and kept fixed).  
On this dataset the model achieves:
- Overall balanced accuracy: **81.09%**
- GPT-3.5 Turbo detection: **100.0%**
- Gemini 2.0 Flash detection: **93.25%**
- Human (arXiv) recall: **75.57%**

## Recommended Citation

```bibtex
@misc{mady2026academictext,
  author       = {Mohamed Mady and Johannes Reschke and Björn Schuller},
  title        = {Academic Text Dataset: A large-scale academic-domain benchmark for AI-generated text detection},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/YOUR_USERNAME/academic-text-detection}},
}
---