LLM Fine-Tuning with PEFT (LoRA)

This project focuses on adapting large language models using Parameter-Efficient Fine-Tuning (PEFT) techniques, specifically LoRA, to specialize a pretrained LLM for a domain-specific task while keeping computational and memory costs low.

I implemented an end-to-end fine-tuning pipeline that includes dataset preparation, tokenizer alignment, LoRA-based adapter training, and evaluation of the adapted model.

Key features:

• Fine-tuning pretrained LLMs using LoRA adapters.
• Efficient training with frozen base weights to reduce GPU memory usage.
• Custom dataset preprocessing and prompt formatting.
• Evaluation of task performance before and after adaptation.
• Support for saving and reloading lightweight adapter weights.

Tech stack: Python, PyTorch, Hugging Face Transformers, PEFT, Datasets, Jupyter Notebook.

Repository: GitHub - fine-tune-llm