83 lines
3.7 KiB
Markdown
83 lines
3.7 KiB
Markdown
---
|
||
license: apache-2.0
|
||
language:
|
||
- ar
|
||
pipeline_tag: text-generation
|
||
tags:
|
||
- 'arabic '
|
||
- text-generation
|
||
widget:
|
||
- text: "أعلنت وزارة الحج في المملكة العربية السعودية"
|
||
example_title: "مثال ١"
|
||
- text: "يبدو اليوم جميلا، سأقوم بتحضير"
|
||
example_title: "مثال ٢"
|
||
- text: "إن التقنيات الحديثة"
|
||
example_title: "مثال ٣"
|
||
---
|
||
# ArabianGPT Model Overview
|
||
|
||
## Disclaimer for the Use of Large Language Models (LLMs) for Text Generation
|
||
|
||
<p style="color: red;">We disclaim all responsibility for any harm, inaccuracies, or inappropriate content generated by ArabianGPT-0.1B, and users engage with and apply the model's outputs at their own risk.</p>
|
||
|
||
> **Important Note:** Currently, we offer a raw pre-trained model. Our team is actively working on releasing instruction-based LLMs that are fine-tuned and augmented with LRHF. The first set of pre-trained models has been made available for community exploration. While we do have models fine-tuned for specific tasks such as summarization and sentiment analysis, they are still in the development phase.
|
||
|
||
|
||
## How you can use this Pre-Trained?
|
||
You are invited to utilize this pre-trained, native Arabic language model as an experimental tool to assess its capabilities, aid in its fine-tuning, and evaluate its performance across a variety of downstream tasks. We encourage you to review our technical report for a comprehensive understanding of the model's performance metrics and the specific downstream tasks it has been tested on. This will provide valuable insights into its applicability and effectiveness in diverse applications.
|
||
|
||
|
||
## Introduction
|
||
ArabianGPT-0.1B, developed under the ArabianLLM initiatives, is a specialized GPT-2 model optimized for Arabic language modeling.
|
||
It's a product of the collaborative efforts at Prince Sultan University's Robotics and Internet of Things Lab, focusing on enhancing natural language modeling and generation in Arabic.
|
||
This model represents a significant stride in LLM research, specifically addressing the linguistic complexities and nuances of the Arabic language.
|
||
|
||
## Key Features
|
||
- **Architecture**: GPT-2
|
||
- **Model Size**: 134 million parameters
|
||
- **Layers**: 12
|
||
- **Model Attention Layers (MAL)**: 12
|
||
- **Context Window Size**: 768 tokens
|
||
|
||
## Training
|
||
- **Dataset**: Scraped Arabic newspaper articles
|
||
- **Data Size**: 15.5 GB
|
||
- **Words**: 237.8 million
|
||
- **Tokenizer**: Aranizer 64K
|
||
- **Tokens**: Over 1.75 billion
|
||
- **Hardware**: 2 NDIVIA A100 GPUs
|
||
- **Training Scale**: 7.5 million examples
|
||
- **Training Duration**: 3 days
|
||
- **Performance**: Final loss of 3.97
|
||
|
||
|
||
## Role in ArabianLLM Initiatives
|
||
ArabianGPT-0.1B (Base Model) is crucial for advancing Arabic language processing, addressing challenges unique to Arabic morphology and dialects.
|
||
|
||
## Usage
|
||
Suitable for Arabic text generation tasks. Example usage with Transformers Pipeline:
|
||
```python
|
||
from transformers import pipeline
|
||
|
||
pipe = pipeline("text-generation", model="riotu-lab/ArabianGPT-01B", max_new_tokens=512)
|
||
text = ''
|
||
pipe.predict(text)
|
||
```
|
||
|
||
## Limitations and Ethical Considerations
|
||
|
||
- The model may have context understanding or text generation limitations in certain scenarios.
|
||
- Emphasis on ethical use to prevent misinformation or harmful content propagation.
|
||
|
||
## Acknowledgments
|
||
|
||
Special thanks to Prince Sultan University, particularly the Robotics and Internet of Things Lab.
|
||
|
||
## Contact Information
|
||
|
||
For inquiries: [riotu@psu.edu.sa](mailto:riotu@psu.edu.sa).
|
||
|
||
## Disclaimer for the Use of Large Language Models (LLMs) for Text Generation
|
||
|
||
<p style="color: red;">We disclaim all responsibility for any harm, inaccuracies, or inappropriate content generated by ArabianGPT-0.1B, and users engage with and apply the model's outputs at their own risk.</p>
|