MohammedNasser/reasoning-silma-1.0-9b

🌟 MohammedNaaser/silma_9b_instruct_ft 🌟

🚀 Fine-Tuned Arabic Language Model for Enhanced Reasoning and Instructive Tasks 🚀

🧠 Model Overview

Welcome to silma_9b_instruct_ft, an advanced fine-tuned version of the SILMA-9B-Instruct, crafted to boost reasoning capabilities and improve response quality in Arabic-language tasks. This model is tailored for causal language modeling (CAUSAL_LM) and has been optimized using parameter-efficient fine-tuning (PEFT) with LoraConfig.

🔧 Key Features

• Reasoning Enhancement 🧩: Fine-tuned specifically for handling complex reasoning tasks in Arabic.
• PEFT for Efficient Fine-Tuning 📊: Employs the LoRA (Low-Rank Adaptation) technique for lightweight and effective fine-tuning.
• Customizable Attention 🎯: Incorporates custom attention modules (q_proj, k_proj, v_proj, etc.) to tailor focus during task processing.

---

🎯 Usage

Here's how to load and start using the model for inference or further fine-tuning:

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("MohammedNaaser/silma_9b_instruct_ft")
model = AutoModelForCausalLM.from_pretrained("MohammedNaaser/silma_9b_instruct_ft")

# Example inference
input_text = """
في إحدى الألعاب، يربح اللاعب 3 نقاط إذا فاز، ويخسر نقطتين إذا خسر
،ولا يحصل على أي نقاط إذا تعادل.
 إذا لعب 20 مباراة وانتهى برصيد 15 نقطة،
 فما أقل عدد من المباريات التي يمكن أن يكون قد فاز بها؟
""""
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🚀 Deployment with Hugging Face Spaces and Gradio

Want to deploy your own interactive chat or Q&A system using this model? Check out Gradio for creating simple UIs!
Create Your Space Here!

---

⚙️ Training Setup

📝 Fine-Tuning Parameters

• Base Model: silma-ai/SILMA-9B-Instruct-v1.0
• Trainer: SFTTrainer from TRL
• Quantization Config: BitsAndBytesConfig (4-bit with nf4 type)
• LoRA Parameters:
  • lora_alpha: 16
  • lora_dropout: 0.1
  • lora_r: 4
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

🛠️ File Overview

After saving the fine-tuned model, the following files are generated:

• adapter_config.json: Configuration file for LoRA adapters.
• adapter_model.safetensor: Serialized fine-tuned model weights.
• tokenizer.json, tokenizer_config.json: Tokenizer configuration files.
• special_tokens_map.json: Mapping for special tokens like [BOS] and [EOS].
• tokenizer.model: The model used for tokenization.

🔑 License & Citation

This model is released under the Apache License. If you use it in your work, please consider citing the original SILMA model and the fine-tuning method as follows:

@misc{MohammedNasser_silma9b_ft,
  author = {Mohammed Nasser},
  title = {Fine-Tuning Arabic Instructive Models with PEFT and LoRA},
  year = {2024},
  url = {https://huggingface.co/MohammedNaaser/silma_9b_instruct_ft},
}

🧑‍💻 Contributing

Contributions are welcome! If you find any issues or would like to improve this model further, feel free to open an issue or submit a pull request. Let's push the boundaries of Arabic NLP together!