---
library_name: transformers
tags:
- legal
datasets:
- Moryjj/Iranian-Courts-Ruling
language:
- fa
metrics:
- rouge
- bertscore
base_model:
- Ahmad/parsT5-base
---

# 🧾 Persian Legal Text Simplification with ParsT5 + Unlimiformer

This model is part of the first benchmark for Persian legal text simplification. It fine-tunes the **ParsT5-base** encoder-decoder model with **Unlimiformer** extension to handle long legal documents efficiently—without truncation.

🔗 **Project GitHub**: [mrjoneidi/Simplification-Legal-Texts](https://github.com/mrjoneidi/Simplification-Legal-Texts)  


---

## 🧠 Model Description

- **Base Model**: [`Ahmad/parsT5-base`](https://huggingface.co/Ahmad/parsT5-base)
- **Extended with**: [Unlimiformer](https://arxiv.org/abs/2309.12254) for long-input attention
- **Training Data**: 5000+ Persian legal rulings, simplified using GPT-4o (ChatGPT) prompts
- **Max Input Length**: ~16,000 tokens (with Unlimiformer)
- **Fine-tuned on**: Simplified dataset generated from judicial rulings
- **Task**: Text simplification (formal → public-friendly)

---

## ✨ Example Usage

```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = AutoModelForSeq2SeqLM.from_pretrained("mrjoneidi/parsT5-legal-simplification")
tokenizer = AutoTokenizer.from_pretrained("mrjoneidi/parsT5-legal-simplification")

input_text = "دادگاه با بررسی مدارک موجود، حکم به رد دعوی صادر می‌نماید..."
inputs = tokenizer(input_text, return_tensors="pt", truncation=True)

output_ids = model.generate(**inputs, max_new_tokens=512)
simplified = tokenizer.decode(output_ids[0], skip_special_tokens=True)

print(simplified)
```


## 🛠 Training Details

- Platform: Kaggle P100 GPU
- Optimizer: AdamW (best performance among AdamW, LAMB, SGD)
- Configurations:
  - 1 vs. 3 unfrozen encoder-decoder blocks
  - Best results with 3-block configuration

---

## More detail on GitHub