Problem. Large language models provide general capabilities, but domain-specific accuracy and tone require targeted fine-tuning. This project demonstrates the workflow for customizing a pretrained model on a curated dataset.

Approach.

• Selected a lightweight pretrained transformer (TinyLlama / GPT-2 variant).
• Preprocessed domain text: cleaning, tokenization, train-test split.
• Used Hugging Face Trainer with transformers + datasets libraries.
• Configured LoRA / PEFT fine-tuning for efficiency on limited GPU.
• Saved intermediate checkpoints and final model weights.
• Logged loss curves and sample generations for qualitative evaluation.

Results (qualitative).

• Model adapted to the dataset’s style and vocabulary after several epochs.
• Fine-tuning improved contextual relevance and reduced generic phrasing.
• Demonstrated parameter-efficient fine-tuning with manageable compute cost.

What I Learned.

• Practical steps for preparing text datasets for language-model training.
• Differences between full fine-tuning and PEFT (LoRA, adapters).
• How hyperparameters (learning rate, batch size, max seq len) affect convergence.
• Managing checkpoints and early stopping in constrained environments.

git clone https://github.com/Joe-Naz01/fine_tune_llm.git
cd fine_tune_llm

conda env create -f environment.yml
conda activate fine_tune_llm
jupyter notebook