Saraa-8B-ORPO-AUNQA

SARAA-8B-ORPO-AUNQA: Self-Assessment Report Analysis Assistant

<div align="center"> <img src="https://img.shields.io/badge/Model-Llama--3--8B-blue" alt="Model"> <img src="https://img.shields.io/badge/License-Apache--2.0-green" alt="License"> <img src="https://img.shields.io/badge/Framework-Unsloth-orange" alt="Framework"> <img src="https://img.shields.io/badge/Specialization-AUN--QA-purple" alt="Specialization"> </div>

📋 Model Description

SARAA-8B-ORPO-AUNQA is a specialized large language model fine-tuned for analyzing Self-Assessment Reports according to ASEAN University Network Quality Assurance (AUN-QA) standards. This model is designed to assist educational institutions in evaluating and improving their quality assurance processes through intelligent document analysis and interactive Q&A capabilities.

• Developed by: StrangeSX
• Model Type: Causal Language Model (Fine-tuned Llama-3-8B)
• Language(s): English, Thai
• License: Apache 2.0
• Finetuned from: unsloth/llama-3-8b-bnb-4bit
• Training Framework: Unsloth 🦥

🎯 Intended Use

Primary Use Cases

• Document Analysis: Analyze self-assessment reports for AUN-QA compliance
• Quality Assurance: Provide insights on educational quality standards
• Interactive Q&A: Answer questions about report content and recommendations
• Educational Assessment: Support institutional evaluation processes

Target Users

• Educational institutions in ASEAN region
• Quality assurance officers
• Academic administrators
• Educational consultants

🚀 Model Performance

🛠️ Technical Details

Training Configuration

• Base Model: Llama-3-8B (4-bit quantized)
• Training Method: ORPO (Odds Ratio Preference Optimization)
• Training Framework: Unsloth + TRL
• Hardware: NVIDIA GPU with 24GB VRAM
• Training Time: ~6 hours (2x faster with Unsloth)
• Memory Usage: 70% less VRAM compared to standard training

Model Architecture

• Parameters: ~8 billion
• Context Length: 8,192 tokens
• Vocabulary Size: 128,256
• Attention Heads: 32
• Hidden Size: 4,096

📊 Training Data

The model was fine-tuned on a curated dataset containing:

• AUN-QA standard documents and guidelines
• Self-assessment report examples
• Quality assurance best practices
• Educational evaluation criteria
• Multi-turn conversation data for Q&A scenarios

🔧 Usage

Quick Start with Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
model_name = "StrangeSX/Saraa-8B-ORPO-AUNQA"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Example usage
prompt = "Analyze this self-assessment report section for AUN-QA compliance:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=512, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Integration with Ollama

# Pull the model
ollama pull strangex/saraa-8b-orpo-aunqa

# Run inference
ollama run strangex/saraa-8b-orpo-aunqa "What are the key criteria for AUN-QA standard 1?"

Web Application Integration

This model is integrated into the SARAA Web Application - a Django-based platform for document analysis:

• Repository: FP_SARAA
• Features: File upload, real-time chat, document vectorization
• Tech Stack: Django, LangChain, ChromaDB, HTMX

⚠️ Limitations and Biases

Known Limitations

• Primarily trained on English and Thai educational documents
• May not generalize well to non-AUN-QA quality standards
• Performance may vary with documents outside the educational domain
• Requires context about AUN-QA standards for optimal performance

Ethical Considerations

• Model outputs should be reviewed by qualified educational professionals
• Not intended to replace human judgment in quality assurance processes
• May reflect biases present in training data

📚 Citation

If you use this model in your research or applications, please cite:

@misc{saraa-8b-orpo-aunqa,
  title={SARAA-8B-ORPO-AUNQA: Self-Assessment Report Analysis Assistant},
  author={StrangeSX},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/StrangeSX/Saraa-8B-ORPO-AUNQA}
}

🔗 Related Resources

• Training Framework: Unsloth - 2x faster LLM training
• Web Application: SARAA Platform
• Base Model: Llama-3-8B
• AUN-QA Standards: Official Documentation

📞 Contact

• Developer: StrangeSX
• GitHub: @StrangeSX

---

This model was trained 2x faster with Unsloth 🦥 and Hugging Face's TRL library.