Are we 75% there?
:D
Good question, also I'm curious as to what the d means?.
I'm not sure if it can be actually considered 75% complete, but I definitely have more data that I'd like to add in the future, as well as train it with 32k tokens-long examples once I'll have the chance.
With the d version I've made some changes (after other unpublished tests) to the data arrangement that should make the model slightly less horny by default, but not added much else that warranted a version bump.
train it with 32k tokens-long examples
Huge.
make the model slightly less horny by default
I am profoundly disappointed and my day is now ruined as a result.
Can you keep the previous version in a v1 branch please?
I haven't removed lewd content—which still composes the majority of the training data—only changed its training order (⇒ non-explicit conversations trained last); that's what I meant with "changes to the data arrangement". From other people's feedback, the model seems paradoxically better in lewd interactions; from personal tests I don't think the behavior changed that much. My plan all along was to take the model onto a more flirty direction rather than sex-obsessed, anyway.
Can you keep the previous version in a v1 branch please?
I'm (obviously) not using git for managing the repository, so that would complicate things up.