3 ms·
I'm sure that's part of it but also in the same way out of distribution tokens can derail its output, certain tokens will guide it. In the paper, the model doe
by famouswaffles 3y ago
I'm sure that's part of it but also in the same way out of distribution tokens can derail its output, certain tokens will guide it.
In the paper, the model doesn't really learn to use padding tokens effectively till after training so it's the only thing.