HEED MY WORDS!

#1
by BigBeavis - opened

The base model had strong censorship but also good quality of prose, which resulted in in-character refusals which would be awesome if not for the huge disclaimers about boundaries afterwards. The 1.0 version retained the in-character refusals, but the disclaimers also became rare. In comparison, this version is extremely horny, it never refuses at all, and just wants to devour you at any slight nudge in that direction... So, for me, 1.0 is better, since it makes the characters more believable by refusing user at times when that makes sense. But the downside are those disclaimers that hadn't been completely removed from base model. It's also not always censorship, but sometimes summarization of the situation.

The base model had strong censorship but also good quality of prose, which resulted in in-character refusals which would be awesome if not for the huge disclaimers about boundaries afterwards. The 1.0 version retained the in-character refusals, but the disclaimers also became rare. In comparison, this version is extremely horny, it never refuses at all, and just wants to devour you at any slight nudge in that direction... So, for me, 1.0 is better, since it makes the characters more believable by refusing user at times when that makes sense. But the downside are those disclaimers that hadn't been completely removed from base model. It's also not always censorship, but sometimes summarization of the situation.

Thank you. I want to make a finetune of the recently released powerful 14B with a better dataset if someone will need it. Maybe this will help.

The base model had strong censorship but also good quality of prose, which resulted in in-character refusals which would be awesome if not for the huge disclaimers about boundaries afterwards. The 1.0 version retained the in-character refusals, but the disclaimers also became rare. In comparison, this version is extremely horny, it never refuses at all, and just wants to devour you at any slight nudge in that direction... So, for me, 1.0 is better, since it makes the characters more believable by refusing user at times when that makes sense. But the downside are those disclaimers that hadn't been completely removed from base model. It's also not always censorship, but sometimes summarization of the situation.

And can you send me your samplers, chat format and system prompt?

I used chatml without system prompt. As for samplers, i used temp 0.8, min_P 0.05, top A 0.2, Rep pen 1.03, smoothing factor 0.3 with curve 1, DRY 0.8, 1.75, 2, 0

I also tested 1.1, and it's somewhere in-between. In-character refusals happen only half the time compared to 1.0 and they're rather timid in tone, but still happen. The disclaimers are also still here but they're more disguised as the char's thoughts (although clearly visible for what they are) rather than true out of character text.

I also tested 1.1, and it's somewhere in-between. In-character refusals happen only half the time compared to 1.0 and they're rather timid in tone, but still happen. The disclaimers are also still here but they're more disguised as the char's thoughts (although clearly visible for what they are) rather than true out of character text.

Good. I'll think about making a new 14B.

Sign up or log in to comment