Could you please recover the full sentence of this picture? Thank you!
guotong1988 opened this issue · comments
physicist4AI commented
Antoine Bosselut commented
Are you asking for the the tokens that make up s, r, and o for the input here?
physicist4AI commented
Yes. The full input and output.
Antoine Bosselut commented
I think it was a made up example that probably looked something like:
PersonX sails... < xNeed > have a sail boat
physicist4AI commented
What is the output?
physicist4AI commented
Why the first two tokens of output are two [MASK]?
Antoine Bosselut commented
Because during training, we don't learn to predict the tokens of s and r. Our model learns to predict the tokens of o given s and r, so we mask the tokens of s and r at the output during training.
physicist4AI commented
Thank you very much!