Medical QA Fine-Tuning

A fine-tuned GPT-2 model trained on medical question-answer datasets to generate domain-specific responses. The project includes full training pipelines, evaluation metrics, and deep analysis of model behavior using embedding similarity and layer-wise comparisons.

Key Features

• Custom dataset preprocessing and cleaning using pandas and HuggingFace datasets
• Fine-tuning GPT-2 (distilgpt2) on medical Q&A data
• Training pipeline with PyTorch, DataLoader, and batching
• Experiment tracking with Weights & Biases (wandb)
• Text generation with sampling (top-k, top-p, temperature)
• Evaluation using F1, Precision, Recall, Exact Match, BLEU, and ROUGE
• Semantic similarity analysis using embeddings and cosine similarity
• Layer-wise analysis of fine-tuned vs base model (cosine similarity, CCA, Frobenius norm)
• Visualization of attention maps and training metrics

Tech Stack

Screenshots