about pre train
#9
by asfjhj - opened
It feels like the pretraining wasn’t sufficient. The model has learned to recognize genitalia, but its understanding of human anatomy seems to have been disrupted and hasn’t been properly rebuilt. Given the current pretraining results, even excellent post-training might not be enough to fix it.
I’m curious how much data the author used for training.
The Base version is better than the RL version with the right prompt / negatives.