---
dataset_name: "KALIMAT"
dataset_summary: "Kalimat - a multipurpose Arabic Corpus containing 18k+ news articles across multiple categories, provided in CSV and JSONL formats."
language:
  - ar
license: "cc-by-4.0"
task_categories:
  - text-generation
  - text-classification
pretty_name: "Kalimat - a multipurpose Arabic Corpus"
size_categories:
  - 10K<n<100K
tags:
  - arabic
  - news
  - corpus
  - nlp
  - low-resource

configs:
  - config_name: csv
    data_files:
      - kalimat.csv

  - config_name: jsonl
    data_files:
      - kalimat.jsonl
---



# Kalimat - a multipurpose Arabic Corpus

This repository provides a cleaned and consolidated version of the **Kalimat - a multipurpose Arabic Corpus**, containing **18,256 Arabic news articles** collected from a diverse range of domains. The original material consisted of thousands of individual `.txt` files organised across multiple category folders. These have been reconstructed, normalised, and compiled into modern machine-learning-friendly formats.

---

## 📚 Corpus Overview

The corpus includes **20k+** articles covering a wide selection of news categories:

- **Politics**
- **Economy**
- **Culture**
- **Religion**
- **Sport**
- **Social / Society-related topics**
- **Other sub-domains depending on the original folder structure**

Each article was originally stored as *one word per line*. In this cleaned edition, all documents have been reconstructed into natural text format with proper spacing, UTF-8 encoding, and consistent metadata extraction.

Although the original filenames varied widely, each document is now associated with the following fields:

- **id** – numeric identifier extracted from the filename (or `-1` where none existed)
- **filename** – original filename exactly as it appeared
- **category** – derived from directory structure or filename
- **year_month** – extracted from filename where possible, otherwise `"unknown"`
- **text** – reconstructed, cleaned article text

---

## 📦 Provided Formats

The cleaned dataset is released in the following forms:

### **1. CSV File**
`kalimat.csv`

A single UTF-8 CSV containing all metadata and article texts.

### **2. JSONL File**
`kalimat.jsonl`

One JSON object per line, suitable for training modern NLP models (e.g. HuggingFace Transformers).

### **3. TXT Version (Zipped)**
`kalimat_txt.zip`

All reconstructed `.txt` documents are included in a single compressed archive to avoid storing thousands of individual files in the repository. Each `.txt` file uses the original filename for easy reference.

---

## 🔀 Train / Validation / Test Splits

The dataset has been randomly split (using a fixed seed for reproducibility) into:

- **Training set** – 80%
- **Validation set** – 10%
- **Test set** – 10%

These splits are provided as:

- `kalimat_train.csv`
- `kalimat_val.csv`
- `kalimat_test.csv`

All CSVs preserve the same column structure as the main file.

### Code used for splitting (for reference)

```
python
import pandas as pd
from sklearn.model_selection import train_test_split

df = pd.read_csv("kalimat.csv", encoding="utf-8")

train_df, temp_df = train_test_split(
    df, test_size=0.20, random_state=42, shuffle=True
)

val_df, test_df = train_test_split(
    temp_df, test_size=0.50, random_state=42, shuffle=True
)

train_df.to_csv("kalimat_train.csv", index=False, encoding="utf-8")
val_df.to_csv("kalimat_val.csv", index=False, encoding="utf-8")
test_df.to_csv("kalimat_test.csv", index=False, encoding="utf-8")
```

### 📁 Repository Structure
```
kalimat.csv
kalimat.jsonl
kalimat_train.csv
kalimat_val.csv
kalimat_test.csv
kalimat_txt.zip
README.md
```

## 📄 Citation

If you use this dataset in your work, please cite the original Kalimat paper:

El-Haj, M., & Koulali, R. (2013). _Kalimat: a multipurpose Arabic corpus_. In Second Workshop on Arabic Corpus Linguistics (WACL-2), pp. 22–25.  
[PDF available here](https://elhaj.uk/docs/KALIMAT_ELHAJ_KOULALI.pdf)  

---