This project fine-tunes Microsoft’s Phi-2 (1.55B parameters) model to:

• Tokenize and preprocess Amazon product data with an optimal max token length of 500
• Use QLoRA to fine-tune the model efficiently, training only ~1.69% (~26M) of total parameters
• Merge fine-tuned weights back into the Phi-2 base model
• Generate high-quality product names and descriptions from Amazon product categories

💡 Optimal data prep with token length tuning for efficient training
⚡ Low-resource fine-tuning using QLoRA with 4-bit quantization
🔄 Merge fine-tuned parameters seamlessly with the base model
🛍️ Generate domain-specific product descriptions and titles

• Model: Microsoft Phi-2 (1.55B params) with QLoRA
• Frameworks: Hugging Face Transformers, PEFT, Accelerate
• Dataset: Custom Amazon product category data
• Training: Efficient tokenization and low-parameter fine-tuning