This model is trained on about 15k lines of a gacha I haven't played in several years. I still have the assets extracting automatically, so I figured to leverage them in order to train an AR model Japanese.

!NOTE!: you are required to set the tokenizer to ./models/tokenizers/japanese.json. Do so under the Settings tab.

This model also has pretty poor zero-shot inferencing capabilities. The voice randomizer is serviceable, as it speaks Japanese pretty decently, but the voice range leaves a lot to be desired.

Use this to quickly finetune a model to any other Japanese speaking character. I've had luck in the past with a different Japanese finetune with just 10 lines.