DocProduct

Why we included this project

DocProduct is a research project, not a product you can drop into a clinical workflow, and the authors are upfront about that. They fine-tuned BERT, starting from BioBERT weights, and GPT-2 on about 700,000 medical question-and-answer pairs scraped from Reddit, HealthTap, and WebMD, then used latent vector search with negative sampling so the model could retrieve relevant medical context before generating a response. For NLP researchers and students, the appeal is a complete pipeline in TensorFlow 2.0, from data preparation through fine-tuning to retrieval and generation, that runs in Colab. The README is explicit that the project should not be used for actionable medical advice, which makes it a solid learning resource and a useful starting point for medical-domain question answering rather than something to deploy.