4 ms·
The model did go from a 100 different trained models to 1. So it seems to hold at least 4x as much in knowledge but maybe we should ask the people that actually
by lowdose 7y ago
The model did go from a 100 different trained models to 1. So it seems to hold at least 4x as much in knowledge but maybe we should ask the people that actually trained it.
Btw thank you for your GPT-2 simple, played around with it last weekend and it made building a toy surprisingly simple!