haikal/haikalbot/ollama_model_recommendations.md
Marwan Alwali 250e0aa7bb update
2025-05-26 15:17:10 +03:00

77 lines
3.2 KiB
Markdown

# Recommended Ollama Models for Arabic Language Support
## Top Recommendations
1. **Jais-13B** (Recommended)
- **Size**: 13 billion parameters
- **Strengths**: Specifically trained on Arabic content, excellent understanding of Arabic context and nuances
- **Command**: `ollama pull jais:13b`
- **Best for**: Production-quality Arabic language understanding and generation
2. **BLOOM-7B**
- **Size**: 7 billion parameters
- **Strengths**: Trained on 46 languages including Arabic, good multilingual capabilities
- **Command**: `ollama pull bloom:7b`
- **Best for**: Multilingual applications where Arabic is one of several languages
3. **Mistral-7B-Instruct**
- **Size**: 7 billion parameters
- **Strengths**: Strong general performance, good instruction following, reasonable Arabic support
- **Command**: `ollama pull mistral:7b-instruct`
- **Best for**: General purpose applications with moderate Arabic requirements
4. **Qwen2-7B**
- **Size**: 7 billion parameters
- **Strengths**: Good multilingual capabilities including Arabic
- **Command**: `ollama pull qwen2:7b`
- **Best for**: Applications requiring both Chinese and Arabic support
## Comparison Table
| Model | Size | Arabic Support | Instruction Following | Resource Requirements | Command |
|-------|------|---------------|----------------------|----------------------|---------|
| Jais-13B | 13B | Excellent | Very Good | High (16GB+ RAM) | `ollama pull jais:13b` |
| BLOOM-7B | 7B | Good | Good | Medium (8GB+ RAM) | `ollama pull bloom:7b` |
| Mistral-7B-Instruct | 7B | Moderate | Excellent | Medium (8GB+ RAM) | `ollama pull mistral:7b-instruct` |
| Qwen2-7B | 7B | Good | Very Good | Medium (8GB+ RAM) | `ollama pull qwen2:7b` |
## Justification for Jais-13B Recommendation
Jais-13B is specifically recommended for your Django AI Analyst application because:
1. **Arabic-First Design**: Unlike most models that treat Arabic as one of many languages, Jais was specifically designed for Arabic language understanding and generation.
2. **Cultural Context**: The model has better understanding of Arabic cultural contexts and nuances, which is important for analyzing domain-specific queries about your data models.
3. **Technical Terminology**: Better handling of technical terms in Arabic, which is crucial for a model analyzing Django models and database structures.
4. **Instruction Following**: Good ability to follow complex instructions in Arabic, which is essential for your prompt-based analysis system.
5. **Performance on Analytical Tasks**: Superior performance on analytical and reasoning tasks in Arabic compared to general multilingual models.
If your system has limited resources (less than 12GB RAM), Mistral-7B-Instruct would be the next best alternative, offering a good balance between performance and resource requirements.
## Installation Instructions
To install the recommended Jais-13B model:
```bash
ollama pull jais:13b
```
For systems with limited resources, install Mistral-7B-Instruct instead:
```bash
ollama pull mistral:7b-instruct
```
After installation, update the `OLLAMA_MODEL` setting in your Django view:
```python
# For Jais-13B
OLLAMA_MODEL = 'jais:13b'
# OR for Mistral-7B-Instruct if resources are limited
# OLLAMA_MODEL = 'mistral:7b-instruct'
```