This repository contains a LoRA adapter fine-tuned on Qwen3-8B to build an Islamic AI assistant named Hadi. The model is designed to be used in RAG (Retrieval-Augmented Generation) systems for providing accurate, context-aware Islamic guidance based on retrieved knowledge.

🧠 Model Overview

Base Model: Qwen3-8B
Fine-tuning Method: LoRA (Low-Rank Adaptation)
LoRA Rank (r): 64
Task: Islamic Question Answering / Assistant behavior tuning
Assistant Name: Hadi
Framework: Hugging Face Transformers + PEFT

📚 Training Dataset

The model was fine-tuned using the following dataset:

👉 Dataset Repository:
Azizkhan22/hadi-islamic-assistant-instruction-dataset

The dataset contains curated Islamic question-answer pairs designed to improve:

Quran and Hadith-based responses
Ethical and daily life guidance in Islamic context
Safe, respectful, and non-extremist responses

⚙️ Intended Use

This adapter is designed for use in RAG pipelines, where external knowledge is retrieved and passed to the model.

Typical Workflow:

Retrieve relevant Islamic knowledge from a vector database (FAISS, pgvector, Pinecone, etc.)
Pass retrieved context + user query to the model
Generate grounded responses using Hadi

🧩 How to Load the Adapter

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "Qwen/Qwen3-8B"

model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto",
    torch_dtype="auto"
)

model = PeftModel.from_pretrained(
    model,
    "YOUR_HF_USERNAME/hadi-qwen3-8b-lora"
)

tokenizer = AutoTokenizer.from_pretrained(base_model_name)

💬 Example System Prompt (Recommended for Inference)

text

You are Hadi, an Islamic assistant.
You provide answers based on authentic Islamic teachings and retrieved context when available.
If context is provided, prioritize it. If not, answer cautiously and avoid speculation.
Always remain respectful and avoid extreme or unsupported claims.

🧠 RAG Integration Notes

This model is designed to work well with retrieval systems. Common embedding models used in similar setups include:

bge-large
e5-large

You can store and retrieve knowledge using:

FAISS
PostgreSQL + pgvector
Pinecone / Weaviate

Always inject retrieved context into the prompt before generation.

⚠️ Limitations

This model is not a replacement for qualified Islamic scholarship
Always verify answers with authentic scholarly sources
May hallucinate if used without proper RAG grounding

📌 Citation

If you use this model, please cite:

Hadi: LoRA-adapted Islamic assistant based on Qwen3-8B using PEFT

❤️ Acknowledgements

Qwen Team for the base model
Hugging Face Transformers & PEFT
Open-source Islamic knowledge resources used in dataset construction

Azizkhan22/qwen3-8b-hadi

Hadi - Islamic Assistant (LoRA Adapter for Qwen3-8B)