Datasets:
Questions in 454k vs 783k without questions
#5
by foss22 - opened
Sprocket-n-Say$ time for f in 000{9..0}.parquet;do arrowcat $f;done|ug -c '\?'
453733
real 0m13,354s
Sprocket-n-Say$ time for f in 000{9..0}.parquet;do arrowcat $f;done|ug -vc '\?'
783289
real 0m13,577s
2 questions or more (chatty):
Sprocket-n-Say$ time for f in 000{9..0}.parquet;do arrowcat $f;done|ug -c '\?\?'
531
The idea is to pretrain chat LM on the chatty half from the beginning, rather then only adding dialogs during SFT stage.
*If it is OK not to train on the bigger half of the data.
For sure I want to train on dialogs as much as possible. If you look at croqaz/vintage-v2 you'll see we have separate dialogs, a lot of books have them separate, thousands of them. I have already seen that the model doesn't know how to chat without lots of dialog examples from the other failed LLM experiments.