IMDB sentiment classification with GPT2 pre training. The main file is here
You can refresh your knowledge about GPT2 from this awesome blog
you can run a similar notebook in colab from here
what is a cls token? it is a token for classification. BERT has this token and it is retrieved from pooling and some other functions such as sigmoid, etc. it is not precisely said in the paper (if you found it please show me so i can edit this). some other folks on github found that it is ok to average each word vector to make and CLS token. so we are adapting to this approach here.
-
a stack overflow to remove html tags from the data set