README.md

---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: label
    dtype: int64
  - name: rater_profile
    sequence: float64
  splits:
  - name: train
    num_bytes: 1198658
    num_examples: 3847
  download_size: 640645
  dataset_size: 1198658
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: cc-by-4.0
task_categories:
- text-classification
language:
- fa
pretty_name: Persian Text Readability Dataset
size_categories:
- 1K<n<10K
---
## Dataset Summary

This is a re-upload of the **Persian Text Readability Dataset**, originally created and published by Mohammadi & Khasteh (2020). It provides sentence-level readability annotations for Persian (Farsi) texts. Each data point includes:

- A **text** in Persian
- A **label** (readability level):  
  - `0` for easy  
  - `1` for medium  
  - `2` for hard
- A **rater profile**: the average readability label distribution of the raters who annotated that specific text

All texts included have over 80% agreement between at least three human annotators. The dataset is intended for training and evaluating readability assessment models in the Persian language.

---

## Supported Tasks and Leaderboards

**Task**: Readability Classification

**Input**: A Persian sentence  
**Output**: A readability level (`0`, `1`, or `2`)

This dataset supports both standard classification models and those that take annotator bias into account (using the `rater_profile`).

---

## Languages

- **Text Language**: `fa` (Persian / Farsi)

---

## Dataset Structure

Each data point is a dictionary with the following fields:

```json
{
  "text": "متن فارسی نمونه",
  "label": 1,
  "rater_profile": [0.1, 0.5, 0.4]
}
```

- `text`: A single Persian sentence or short passage.
- `label`: The final readability level (`0`: easy, `1`: medium, `2`: hard).
- `rater_profile`: A 3-element float list showing the average readability preferences of the annotators.

### Dataset Stats

| **Level** | # of texts | Avg. words per text |
|----------|------------|---------------------|
| 0 (easy) | 2,953      | 28.8                |
| 1 (medium) | 572      | 39.8                |
| 2 (hard) | 322        | 62.1                |
| **Total** | **3,847** | **33.2**            |

---

## Source Data

### Annotation Process

Texts were manually rated by undergraduate students at the **K. N. Toosi University of Technology**. Each text was rated by at least three annotators. Only texts where at least 80% of the raters agreed on the label were included.

### Rater Profile

The `rater_profile` field helps capture rater bias. For example, `[0.1, 0.5, 0.4]` means the raters of that text tend to give:
- 10% of their scores as "easy"
- 50% as "medium"
- 40% as "hard"

This can be useful in modeling subjective readability with annotator-specific information.

---

## Citation

Please cite the following if you use this dataset:

```bibtex
@inproceedings{mohammadi2020machine,
  title={A machine learning approach to Persian text readability assessment using a crowdsourced dataset},
  author={Mohammadi, Hamid and Khasteh, Seyed Hossein},
  booktitle={2020 28th Iranian Conference on Electrical Engineering (ICEE)},
  pages={1--7},
  year={2020},
  organization={IEEE}
}
```

---

## Acknowledgements

We express deep appreciation to the undergraduate computer engineering students at the **K. N. Toosi University of Technology** who annotated the dataset.

---

## Licensing

This dataset is a re-upload. Licensing terms are inherited from the original work. Please ensure compliance with any applicable usage conditions described in the original publication or source repository: [https://github.com/sandstorm12/persian_readability_dataset](https://github.com/sandstorm12/persian_readability_dataset)