Medical-Assistant

• The given dataset had two columns for question and answer. The same questions were repeated multiple times with different answers, so first, similar questions were aggregated and combined the answers for them to get a common context out of them.

• Used the given dataset to train the BART model and tokenizer.

• I also tried with T5 but gave a meager score. Can't use bert as it works well on a dataset given context and question as input to find an answer in that context.

• This model gave training and validation loss as:

  

• For training the model, I have used a batch size of 16 for both datasets and epochs, which are set to be 5; more than that would have cost me more computational power and time.

• I could have used the F1 score and exact match for this model's evaluation.

  Exact Match: It indicates how many answers were correct. It helps understand the model's ability to generate precise and accurate responses.

  F1-Score: It balances precision and recall to show how well the model captures the relevant information in its predictions. It prioritizes the metrics that is less in value.

• Used this model to predict on some medical-based questions:

  

Improvements:

• Could have used a better dataset to train the model.
• A squad v2 model is also trained on medical articles for question-and-answer but uses contents apart from the questions as input for answering the questions.
• Could have trained for much longer as the curve converged when the training was stopped.