Comment by gwern
11 days ago
Cite? I don't see how either of those could deal with the fact that the logits become uninformative and 'flattened' after the tuning. How can a sampler undo the erasure of information?
11 days ago
Cite? I don't see how either of those could deal with the fact that the logits become uninformative and 'flattened' after the tuning. How can a sampler undo the erasure of information?
No comments yet
Contribute on Hacker News ↗