amitca71/farsi_syntetic_openai4omini_vs_real

Persian Real vs. Synthetic Text Dataset (PRSD)

This dataset is a consolidated benchmark for Persian (Farsi) LLM detection. It combines human-authored text from established Persian NLP corpora with high-quality synthetic text generated by OpenAI's GPT-4o-mini.

The dataset is pre-shuffled and split into Train, Validation, and Test sets, designed specifically for training binary classifiers to distinguish between human and machine-generated Persian text.

---

Dataset Composition

The dataset consists of 33,747 samples balanced between two distinct classes:

Class 0: Human-Authored Text (Real)

Collected from three major Persian NLP datasets, representing a mix of formal, informal, and handwriting-transcribed text:

1. Parsoff: Offline handwriting recognition dataset (short, varied text).
2. DeepSentiPers: Sentiment analysis corpus (social media, informal, slang).
3. ParsiNLU: Natural Language Understanding benchmark (formal, literary, news).

Class 1: Synthetic Text (Artificial)

Generated using OpenAI GPT-4o-mini (2024–2025).

• Method: Prompt-based generation designed to stylistically mimic the categories above (formal/informal) without directly paraphrasing or plagiarizing the original samples.
• Goal: To create a challenging adversary for detection models by simulating high-quality, fluent Persian text.

---

Task Definition

This is a Binary Classification task:

• 0 (Negative Class): Real / Human
• 1 (Positive Class): Fake / Synthetic

---

Dataset Structure

Data Fields

• text (string): The input text sample (Persian).
• label (int64):
  • 0: Human-authored.
  • 1: AI-generated.

Splits

Split	Examples	Percentage
Train	26,997	~80%
Validation	3,375	~10%
Test	3,375	~10%

Splits were created using stratified sampling to ensure the ratio of Real vs. Synthetic text remains consistent across all partitions.

---

Intended Use

This dataset is intended for Non-Commercial Academic Research Only:

• LLM Detection: Training classifiers to detect AI-generated Persian text.
• Stylometry: Analyzing linguistic differences between human and machine Farsi.
• Robustness Testing: Evaluating how well models generalize across different domains (e.g., social media vs. formal writing).

Prohibited Use

• ❌ Commercial applications.
• ❌ Surveillance or authorship profiling of individuals.
• ❌ Use in production content moderation systems without human oversight.

---

Ethical Considerations & Privacy

1. Human Data (Class 0)

• Source: Publicly available datasets (Parsoff, DeepSentiPers, ParsiNLU).
• Privacy Warning: While the original datasets were anonymized, text derived from social media or handwriting samples may still contain Personally Identifiable Information (PII).
• Biometric Data (Parsoff): The Parsoff subset includes transcriptions of handwriting samples. Handwriting patterns are considered biometric identifiers. While this dataset contains only the text (not the images), users should be aware of the sensitive nature of the original source.
• Responsibility: Users must comply with GDPR and ethical guidelines when handling the human subset.

2. Synthetic Data (Class 1)

• Source: Generated by AI (GPT-4o-mini).
• Bias: Reflects the biases inherent in the generating model.
• Nature: Fully artificial; any resemblance to real events or persons is coincidental.

---

License

This dataset is released as a composite work. The aggregate dataset is licensed under CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International).

Why this license? This restrictive license is required because the source datasets (specifically ParsiNLU) carry a ShareAlike clause. By using this dataset, you agree to:

1. Attribution: Cite both the original dataset authors (listed below) and this repository.
2. Non-Commercial: Use this data only for research purposes.
3. ShareAlike: If you remix or build upon this data, you must distribute your contributions under the same license.

Fair Use & Redistribution: This repository hosts a transformed subset of the original corpora, curated specifically for the purpose of benchmarking synthetic text detection. It is not intended to serve as a mirror or replacement for the original full datasets. We encourage researchers to download the full original datasets from their respective official sources for other tasks.

---

Credits & Citations

You must cite the original authors of the human subsets when using this dataset.

1. Human Data Sources

ParsiNLU

@article{khashabi2020parsinlu,
  title={ParsiNLU: A Suite of Language Understanding Datasets for Persian},
  author={Khashabi, Daniel et al.},
  journal={arXiv preprint arXiv:2012.06154},
  year={2020}
}

@article{sharami2020deepsentipers,
  title={DeepSentiPers: Novel Deep Learning Models Trained Over Proposed Augmented Persian Sentiment Corpus},
  author={Sharami, Javad PourMostafa Roshan and Sarabestani, Parsa Abbasi and Mirroshandel, Seyed Abolghasem},
  journal={arXiv preprint arXiv:2004.05328},
  year={2020}
}
@article{ebrahimi2022parsoff,
  title={Parsoff: A specific dataset for offline handwritten Persian recognition},
  author={Ebrahimi, M. and Torkaman, A.},
  journal={Journal of AI and Data Mining},
  volume={10},
  number={1},
  pages={111--120},
  year={2022}
}