Training Wiki103 From Scratch


(Bobak Farzin) #1

I am curious to see the notebook or codeblock that generated the pre-trained weights for the Wiki103 data set using either FastAI V1 or an earlier version.

I am trying to duplicate that process to improve my understanding and I am struggling with what params were used, how much training, what was loaded as the tokenizer, what the corpus looked like. Does someone have that code and are they willing to share it so I can better understand where the pre-trained weights come from and what a “good” model looks like in terms of training path and params?