Author - Hayden Beadles
This model is meant to evaluate the results of creating an Encoder / Decoder generative model using SciBERT. The model is then finetuned on 30000 samples of the PubMedQA dataset. Instead of being finetuned
on the columns
question
and
final_answer
, where
final_answer
is a set of yes / no answers, we instead fine tune on the more challenging
long_answer
column, which gives a short answer
to the question.
The model was fine-tuned over 3 epochs, using the Adam learning rate scheduler, with a max length of 128 tokens.
The results are to help gauge SciBERT's abilities to answer (generate an answer) directly to a question, with no context provided. It is meant to evaluate the overall models training and attention towards
a more focused topic, to see if SciBERTs base training gives it any advantages.