AliAbdelrasheed/maqa_llama

MAQA-LLaMA — Arabic Medical Question Answering (Base Model)

⚠️ Disclaimer: This model is intended for research and informational purposes only. It is not a substitute for professional medical advice, diagnosis, or treatment. It cannot and should not be used to prescribe or recommend medications.

---

Model Summary

maqa_llama is a Meta Llama 3 8B Instruct model fine-tuned on the MAQA dataset (Medical Arabic Questions & Answers) — 430,000 real doctor-patient interactions across 20 medical specialisations, sourced from Arabic medical platforms.

This is the full-precision base model (BF16 / 16-bit), best suited for research and further fine-tuning. For deployment on consumer hardware, see the quantised variants below.

Property	Value
Base model	unsloth/llama-3-8b-Instruct-bnb-4bit (Meta Llama 3 8B Instruct)
Fine-tuning method	QLoRA (via Unsloth)
Precision	BF16 (merged 16-bit)
Model size	8B parameters
Language	Arabic 🇸🇦
License	Apache 2.0
Developed by	Ali Abdelrasheed

---

Model Family

Model	Format	Size	Best for
maqa_llama ← this model	BF16 SafeTensors	~16 GB	Research / further fine-tuning
maqa_llama_4bit	4-bit (bitsandbytes)	~5 GB	GPU inference
maqa_llama_4bit_GGUF	GGUF q4_k_m	4.92 GB	CPU / local deployment

---

Dataset — MAQA

The model was trained on MAQA (Medical Arabic Questions & Answers), the largest Arabic medical Q&A dataset available for NLP research.

Property	Value
Total records	430,000 question-answer pairs
Columns	question (patient) · answer (doctor diagnosis + treatment notes)
Medical specialisations	20 (cardiology, dermatology, neurology, gastroenterology, paediatrics, and more)
Sources	altibbi.com · tbeeb.net · cura.healthcare
Language	Arabic (Modern Standard Arabic)
Quality	All questions unique and cleaned (not stemmed)
Training split	70% train / 30% evaluation

Dataset reference: "Deep learning for Arabic healthcare: MedicalBot" — Social Network Analysis and Mining, Springer (2023) Available on Harvard Dataverse

---

Training Details

LoRA Configuration

Hyperparameter	Value
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0 (optimised)
Bias	none (optimised)
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Gradient checkpointing	unsloth (30% less VRAM)
Random state	3407

Training Arguments

Hyperparameter	Value
Epochs	1
Batch size (per device)	52
Gradient accumulation steps	1
Learning rate	2e-4
LR scheduler	Linear
Warmup steps	200
Optimiser	AdamW 8-bit
Max sequence length	2048
Evaluation strategy	Every 500 steps
Training environment	Google Colab Pro
Framework	Unsloth + HuggingFace TRL (SFTTrainer)

Chat Template & System Prompt

The model was fine-tuned using the Llama-3 chat template with a custom Arabic medical system prompt:

أنت طبيب محترف ولديك خبرة في كل مجالات الطب.
يجيب على أسئلة المرضى حول الأمراض، باستخدام لهجة رسمية وودية،
وإجابات موجزة ومفيدة يسهل على الجميع فهمها.

"You are a professional doctor with expertise in all fields of medicine. Answer patients' questions about diseases using a formal and friendly tone, with concise and helpful answers that everyone can understand."

Special tokens <|question|> and <|answer|> were added to the tokenizer to clearly demarcate patient input and doctor response during training.

---

Training Journey

The development of this model went through multiple iterations:

1. RAG approach (initial attempt): Explored Retrieval-Augmented Generation using LangChain and FAISS as a vector store. While technically functional, this approach had limitations in generalisation and was not suitable for a fine-tuned conversational model.

2. Manual LoRA fine-tuning: Moved to direct supervised fine-tuning using LoRA for full control over the model weights on the MAQA dataset.

3. Unsloth optimisation (final): Adopted the Unsloth framework to maximise training efficiency on Colab Pro's GPU resources — achieving 2x faster training with significantly reduced VRAM usage. This produced the current published models.

---

Quick Start

from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "AliAbdelrasheed/maqa_llama",
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = False,  # Full precision
)

tokenizer = get_chat_template(tokenizer, chat_template="llama-3")
FastLanguageModel.for_inference(model)

messages = [
    {"from": "system", "value": "أنت طبيب محترف ولديك خبرة في كل مجالات الطب. يجيب على أسئلة المرضى حول الأمراض، باستخدام لهجة رسمية وودية، وإجابات موجزة ومفيدة يسهل على الجميع فهمها."},
    {"from": "human", "value": "ما هي أعراض مرض السكري من النوع الثاني؟"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

from transformers import TextStreamer
streamer = TextStreamer(tokenizer)
_ = model.generate(input_ids=inputs, streamer=streamer, max_new_tokens=256, use_cache=True)

---

Limitations

• Not a substitute for professional medical advice or clinical diagnosis
• Cannot prescribe, recommend, or endorse specific medications
• Trained on a sampled subset of MAQA — performance may vary across all 20 specialisations
• Optimised for Modern Standard Arabic; dialectal Arabic (Egyptian, Levantine, etc.) performance may vary
• Web-scraped training data may contain noise or outdated medical information
• Scarcity of high-quality Arabic medical data remains an open challenge in the field

---

Future Work

• Fine-tuning on the full 430,000-row MAQA dataset (current model trained on a sampled subset)
• Multi-turn conversational memory for sustained patient-doctor dialogue
• Dialect-specific fine-tuning (Egyptian Arabic, Gulf Arabic)
• Multimodal input support (e.g. dermatology image input)
• Integration with electronic medical records (EMR) systems
• Real-time knowledge updates from medical literature

---

Citation

If you use this model in your research, please cite the MAQA dataset:

@article{maqa2023,
  title={Deep learning for Arabic healthcare: MedicalBot},
  journal={Social Network Analysis and Mining},
  publisher={Springer},
  year={2023}
}

---

Developed By

Ali Abdelrasheed — Graduation Project Nile University · B.Sc. Information Technology – Big Data · Class of 2024 🤗 HuggingFace Profile