|
--- |
|
language: |
|
- fa |
|
metrics: |
|
- bertscore |
|
- rouge |
|
base_model: |
|
- Ahmad/parsT5-base |
|
pipeline_tag: text-generation |
|
tags: |
|
- legal |
|
- Simplification |
|
- text-to-text |
|
--- |
|
# Persian Simplification Model (parsT5 Base) |
|
|
|
--- |
|
|
|
## Overview |
|
|
|
This model is a fine-tuned ParsT5 (base) version designed explicitly for the Persian Simplification Task. The training data consists of Persian legal texts. The model is trained using supervised fine-tuning and employs the **Unlimiformer Algorithm** to handle large inputs effectively. |
|
|
|
- **Architecture**: Ahmad/parsT5-base |
|
- **Language**: Persian |
|
- **Task**: Text Simplification |
|
- **Training Setup**: |
|
- **Algorithm for reducing computation**: Unlimiformer |
|
- **Epochs**: 12 |
|
- **Hardware**: NVIDIA GPU 4070 |
|
- **Trainable Blocks**: Last Encoder-Decoder |
|
- **Optimizer** : AdamW + lr_scheduler |
|
- **Input max Tokens**: 4096 |
|
- **Output max Tokens**: 512 |
|
|
|
--- |
|
|
|
## Readability Scores |
|
|
|
The following table summarizes the readability scores for the original texts and the predictions generated by the fine-tuned model: |
|
|
|
| Metric | Original Texts | Predictions | |
|
|----------------|----------------|-------------| |
|
| Gunning Fox | 14.9676 | **7.5891** | |
|
| ARI | 11.8796 | **6.7869** | |
|
| Dale-Chall | 2.6473 | **1.2679** | |
|
| Flesch-Dayani | 228.2377 | **244.0153**| |
|
|
|
--- |
|
|
|
## Evaluation Results |
|
|
|
The fine-tuned model was evaluated using **Rouge** and **BERTScore (mBERT)** metrics. For comparison, the performance of two other Persian LLMs based on LLaMA is also presented: |
|
|
|
|
|
| Prediction Model | Rouge1 | Rouge2 | RougeL | Precision | Recall | F1 | |
|
|-----------------------------------------------|---------|---------|---------|-----------|---------|---------| |
|
| Fine-Tuned Model | **38.08%** | **15.83%** | **19.41%** | **76.75%** | 71.06% | **73.71%** | |
|
| ViraIntelligentDataMining/PersianLLaMA-13B | 28.64% | 9.81% | 13.67% | 68.36% | 73.44% | 70.80% | |
|
| MehdiHosseiniMoghadam_AVA_Llama_3_V2 | 30.07% | 10.33% | 16.39% | 68.47% | **73.47%** | 70.87% | |
|
|
|
--- |
|
|
|
## How to Use |
|
|
|
You can load and use this model with the Hugging Face library as follows: |
|
|
|
```python |
|
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer |
|
|
|
# Load the model |
|
tokenizer = AutoTokenizer.from_pretrained("Moryjj/FineTuned-parsT5-Simplification") |
|
model = AutoModelForSeq2SeqLM.from_pretrained("Moryjj/FineTuned-parsT5-Simplification") |
|
|
|
# Example usage |
|
input_text = "متن پیچیده فارسی" |
|
inputs = tokenizer(input_text, return_tensors="pt", max_length=4096, truncation=True) |
|
outputs = model.generate(**inputs) |
|
|
|
# Decode the output |
|
simplified_text = tokenizer.decode(outputs[0], skip_special_tokens=True) |
|
print(simplified_text) |
|
``` |
|
|
|
### Contact Information |
|
For inquiries or feedback, please contact: |
|
|
|
Author: Mohammadreza Joneidi Jafari |
|
|
|
Email: [email protected] |
