This model is a fine-tuned version of Meta's Llama-3.2-1B, optimized specifically for the Arabic language. It was trained using the Alpaca instruction dataset, enabling it to follow user commands, answer questions, and generate creative text in Arabic with improved linguistic flow.

🛠 Model Description

Developed by: shamil88
Model Type: Causal Language Model
Language(s): Arabic (Primary), English (Secondary)
Base Model: Meta-Llama-3.2-1B
Fine-tuning Technique: LoRA (Low-Rank Adaptation)
Quantization: 4-bit (bitsandbytes) for training, GGUF (Q4_K_M) for deployment.

🚀 Technical Specifications

Training Environment

The model was trained on a local machine using the following setup:

GPU: NVIDIA Quadro T2000 (4GB VRAM)
Platform: WSL2 (Ubuntu 22.04)
Optimization: Unsloth AI (2x faster training and 70% less memory usage).

Hyperparameters

Parameter	Value
Steps	100
Max Sequence Length	512
Learning Rate	2e-4
Batch Size	2
Gradient Accumulation	4
Optimizer	AdamW (8-bit)
Scheduler	Linear

📈 Training Results

Starting Loss: ~1.88
Ending Loss: ~1.38 The significant drop in loss within 100 steps indicates that the model successfully adapted to the Arabic instruction-response patterns provided in the dataset.

📦 Quantization & Deployment

This repository contains:

LoRA Adapters: Lightweight files (adapter_model.bin) to be used with the base model.
GGUF Files: Pre-quantized files (q4_k_m) compatible with LM Studio, llama.cpp, and Ollama, making it easy to run on consumer-grade hardware or mobile devices.

💻 How to Use

Using Unsloth (Recommended for Inference)

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "shamil88/Arabic-Llama-3.2-1B-Alpaca",
    max_seq_length = 512,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# Your Arabic Prompt
prompt = "### Instruction:\nما هي فوائد القراءة؟\n\n### Response:\n"
inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens = 128)
print(tokenizer.batch_decode(outputs))