nice!
@juiceb0xc0de What training pipeline did you use to put the personality in? We may be able to adapt it to put reasoning in G1 later lol
I did a selective weight freezing on a few compliance features and did a full fine-tune. Gemma was one of the hardest models I ever encountered while trying to train with a personality so the selective freezing was important. I also used a custom data set I made awhile back. I took conversations I had with different frontier models and flipped the roles. So I was the assistant and the model was the user. Then I trained on response only. Its 100% just my voice. I also grabbed any texts I could find and i ended up with just over 10k rows. It's a fun way to learn about training. And super easy to know if you were successful or not.
Honestly, I think it's because the model has bartender in the title so people think its like talking to the bartenders in skyrim sort of thing. I legit had a user save it to a NPC's for skyrim collection. I don't know if i ever once posted about the models on any platform. I just make them for my own curiosity and entertainment. And its all my dialogue, I'm a bartender so I just called it a bartender since that's what it's based off of.