---
library_name: transformers
tags:
- Legal
- court
- prediction
- Arabic
- NLP
datasets:
- mbayan/Arabic-LJP
language:
- ar
base_model:
- meta-llama/Llama-3.1-8B-Instruct
pipeline_tag: text-generation
---

# Arabic Legal Judgment Prediction Dataset

## Overview

This dataset is designed for **Arabic Legal Judgment Prediction (LJP)**, collected and preprocessed from **Saudi commercial court judgments**. It serves as a benchmark for evaluating Large Language Models (LLMs) in the legal domain, particularly in low-resource settings.

The dataset is released as part of our research:

> **Can Large Language Models Predict the Outcome of Judicial Decisions?**  
> *Mohamed Bayan Kmainasi, Ali Ezzat Shahroor, and Amani Al-Ghraibah*  
> [arXiv:2501.09768](https://arxiv.org/abs/2501.09768)
##  Model Usage
```python

To use the model
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mbayan/Llama-3.1-8b-ArLJP")
model = AutoModelForCausalLM.from_pretrained("mbayan/Llama-3.1-8b-ArLJP")
```

## Dataset Details

- **Size:** 3752 training samples, 538 test samples.
- **Annotations:** 75 diverse Arabic instructions generated using GPT-4o, varying in length and complexity.
- **Tasks Supported:**  
  - Zero-shot, One-shot, and Fine-tuning evaluation of Arabic legal text understanding.

## Data Structure

The dataset is provided in a structured format:

```python
from datasets import load_dataset

dataset = load_dataset("mbayan/Arabic-LJP")
print(dataset)
```

The dataset contains:
- **train**: Training set with 3752 samples
- **test**: Test set with 538 samples

Each sample includes:
- **Input text:** Legal case description
- **Target text:** Judicial decision

## Benchmark Results

We evaluated the dataset using **LLaMA-based models** with different configurations. Below is a summary of our findings:

| **Metric**               | **LLaMA-3.2-3B** | **LLaMA-3.1-8B** | **LLaMA-3.2-3B-1S** | **LLaMA-3.2-3B-FT** | **LLaMA-3.1-8B-FT** |
|--------------------------|------------------|------------------|---------------------|---------------------|---------------------|
| **Coherence**             | 2.69             | 5.49             | 4.52                | *6.60*              | **6.94**            |
| **Brevity**               | 1.99             | 4.30             | 3.76                | *5.87*              | **6.27**            |
| **Legal Language**        | 3.66             | 6.69             | 5.18                | *7.48*              | **7.73**            |
| **Faithfulness**          | 3.00             | 5.99             | 4.00                | *6.08*              | **6.42**            |
| **Clarity**               | 2.90             | 5.79             | 4.99                | *7.90*              | **8.17**            |
| **Consistency**           | 3.04             | 5.93             | 5.14                | *8.47*              | **8.65**            |
| **Avg. Qualitative Score**| 3.01             | 5.89             | 4.66                | *7.13*              | **7.44**            |
| **ROUGE-1**               | 0.08             | 0.12             | 0.29                | *0.50*              | **0.53**            |
| **ROUGE-2**               | 0.02             | 0.04             | 0.19                | *0.39*              | **0.41**            |
| **BLEU**                  | 0.01             | 0.02             | 0.11                | *0.24*              | **0.26**            |
| **BERT**                  | 0.54             | 0.58             | 0.64                | *0.74*              | **0.76**            |

**Caption**: A comparative analysis of performance across different LLaMA models. The model names have been abbreviated for simplicity: **LLaMA-3.2-3B-Instruct** is represented as LLaMA-3.2-3B, **LLaMA-3.1-8B-Instruct** as LLaMA-3.1-8B, **LLaMA-3.2-3B-Instruct-1-Shot** as LLaMA-3.2-3B-1S, **LLaMA-3.2-3B-Instruct-Finetuned** as LLaMA-3.2-3B-FT, and **LLaMA-3.1-8B-Finetuned** as LLaMA-3.1-8B-FT.

### **Key Findings**
- Fine-tuned smaller models (**LLaMA-3.2-3B-FT**) achieve performance **comparable to larger models** (LLaMA-3.1-8B).
- Instruction-tuned models with one-shot prompting (LLaMA-3.2-3B-1S) significantly improve over zero-shot settings.
- Fine-tuning leads to a noticeable boost in **coherence, clarity, and faithfulness** of predictions.

## Usage

To use the dataset in your research, load it as follows:

```python
from datasets import load_dataset

dataset = load_dataset("mbayan/Arabic-LJP")

# Access train and test splits
train_data = dataset["train"]
test_data = dataset["test"]
```

## Repository & Implementation

The full implementation, including preprocessing scripts and model training code, is available in our GitHub repository:

🔗 **[GitHub](https://github.com/MohamedBayan/Arabic-Legal-Judgment-Prediction)**

## Citation

If you use this dataset, please cite our work:

```
@misc{kmainasi2025largelanguagemodelspredict,
  title={Can Large Language Models Predict the Outcome of Judicial Decisions?}, 
  author={Mohamed Bayan Kmainasi and Ali Ezzat Shahroor and Amani Al-Ghraibah},
  year={2025},
  eprint={2501.09768},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2501.09768}, 
}
```