Closed redthing1 closed 3 years ago
It's good for objective metrics (which is what the language modeling research papers use), but in the end for text generation, there's a lot of subjectivity (especially with temperatures) which a validation pass don't help as much to identify.
I was curious about this in your features page:
Could you explain a bit? I was under the impression that validation was a useful metric for training causal language models.