3 ms·
It's the same process as regular training, but instead of a single token for your cost function, it's top-k logits.
by qeternity 1y ago
It's the same process as regular training, but instead of a single token for your cost function, it's top-k logits.