---
language:
- ar
- en
tags:
- chat
- translation
license: apache-2.0
base_model: OMDA-Decoder
pipeline_tag: text-generation
---

# OMDA: Bilingual Arabic-English Chat LLM

**Model Name:** OMDA  
**Architecture:** OMDA-Decoder (Custom Architecture)  
**Tokenizer:** OMDATokenizer (Custom Tokenizer)  
**Languages:** Arabic (Primary), English  
**Model Type:** Chat/Instruction-following  
**Author:** Binomda  
**Release Date:** 2025-06-28  

## Model Overview
OMDA is a compact bilingual language model specifically designed for Arabic-English conversational AI applications. Built with a custom decoder architecture, it excels at understanding and generating natural responses in both languages.

## Model Specifications
| Parameter          | Value     |
|--------------------|-----------|
| Layers             | 6         |
| Hidden Size        | 512       |
| Attention Heads    | 8         |
| FFN Dimension      | 2048      |
| Max Sequence Length| 512       |
| Vocabulary Size    | 128,004   |
| Training Data      | 1,000 curated Arabic-English conversation pairs |

## Intended Uses
✅ Chatbot development  
✅ Bilingual assistant applications  
✅ Educational tools  
✅ Basic translation tasks  

## Quick Start
```python
from transformers import pipeline

# Initialize chat pipeline
chatbot = pipeline("text-generation", model="BINOMDA/OMDA")

# Arabic input example
ar_responser = chatbot("ما هو رأيك في التكنولوجيا الحديثة؟")

# English input example
en_response = chatbot("Explain artificial intelligence simply")

## Intended Use
- Chatbots, assistants, translation, and educational tools for Arabic/English.

## Training
- Trained for 5 epochs on 1000 samples.
- Loss curve and checkpoints included.

## Limitations
- This is a small-scale demonstration model and may not generalize well to all real-world chat scenarios.
- Not suitable for production use without further scaling, extensive evaluation, and safety checks.
- Limited training data and model size may result in hallucinations or inaccurate translations.
- No advanced filtering for inappropriate or biased outputs.
- For research and educational purposes only.

## Export & Deployment
- See below for HuggingFace, llama.cpp, and ollama export