3 ms·
I mean, this depends on your sampler. With temp=1 and sampling from the raw output distribution, setting aside numerics issues, these models output nonzero prob
by anonymoushn 1y ago
I mean, this depends on your sampler. With temp=1 and sampling from the raw output distribution, setting aside numerics issues, these models output nonzero probability of every token at each position
- danielmarkbruce 1y agoA large model well trained on good data will have logits so negative for something like "1+1=" -> 3 that they won't come up in practice unless you sample in a way to deliberately misuse the model.