Nice 🚀 Adding a clear README.md to your Hugging Face repo will make Ibn-Al-Nafs stand out and help others understand your model quickly. Here’s a polished draft you can use (you can copy it into README.md):

---

Ibn-Al-Nafs 🕌📖

Ibn-Al-Nafs is a retrieval-augmented fine-tuned model for Arabic cultural multiple-choice question answering (MCQs), developed as part of the PalmX 2025 Shared Task. It adapts NileChat-3B using parameter-efficient fine-tuning (PEFT) and integrates Gemini-based retrieval to provide culturally grounded evidence during training.

🏆 Leaderboard Result: Ranked 6th place on PalmX 2025 Subtask 1 with 67.6% accuracy, outperforming the NileChat-3B baseline by +3%.

📄 Associated Paper: ISL-NLP at PalmX 2025: Retrieval-Augmented Fine-Tuning for Arabic Cultural Question Answering (Accepted, 2025).

---

Model Details

• Base Model: NileChat-3B

• Techniques:

  • Retrieval-Augmented Training with Gemini API
  • Parameter-Efficient Fine-Tuning (PEFT): updating only q_proj, v_proj, and gate_proj layers (68.2% reduction in trainable params)
  • Instruction-style fine-tuning in Modern Standard Arabic

---

Results

On the PalmX 2025 Subtask 1 (Arabic Cultural MCQs) development set, Ibn-Al-Nafs outperforms both the NileChat-3B baseline and general-purpose Arabic LLMs:

Model	Precision	Recall	F1-score	Accuracy
Qwen2.5-1.5B	64.73	63.89	63.59	63.60
Qwen1.5-1.8B	63.24	60.88	59.15	59.80
NileChat-3B	71.74	70.00	69.92	70.00
Ibn-Al-Nafs	73.81	73.88	73.54	73.60

🏆 Ranked 6th place on the official PalmX 2025 leaderboard, achieving 67.6% accuracy on the blind test set.

---

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "MohamedGomaa30/Ibn-Al-Nafs"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Example Arabic cultural MCQ
question = "من هو الطبيب الشهير الذي كتب كتاب القانون في الطب؟"
options = ["ابن رشد", "ابن سينا", "الفارابي", "الزهراوي"]

inputs = tokenizer(f"السؤال: {question}\nالاختيارات: {', '.join(options)}\nالإجابة:", return_tensors="pt")
outputs = model.generate(**inputs, max_length=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Acknowledgments

• Developed at Intelligent Systems Lab (ISL-NLP), Arab Academy for Science & Technology.
• Thanks to the PalmX 2025 organizers and Kaggle for resources.

---