25%
Hi there! Thank you for releasing an open source model. I saw that you were looking for opinions. I tested a few of your modes, including the newer v2 version. In my opinion, the 25% v1.1 is by far the best. V2 seems to have a very hard time with coherency. The 25% has no problems with coherency and is fun to use, I find it to be the perfect mix!
Ps, maybe add anime dataset in the model in the future? :)
Thank you! I'll definitely broaden the scope once I have a good handle on finetuning. (In fact, let me add a few more anime content before starting with v3)
I appreciate all the feedback, including your v2 tests. Seems like I'll need to hold back on the moist to maintain coherency, but I hope you've still had a unique moist experience even with 25%.
Looking forward to more feedback like this :D
Can I ask you what settings and instructions you used to do the tests? It’s just that in my case, all the models most often gave interesting answers, but unfortunately they were just as often not logical or... they didn’t accurately follow what was written on the character’s card.
Sure! These are my settings on SillyTavern.
Hosted on colab, btw.
Context: 8192
Temp: 0.8
Top K: 0
Top P: 0.95
Typical P: 1
Min P: 0
Top A: 0
Tail Free Sampling: 1
Repetition Penalty: 1.1
Roleplay context template+Story string (I don't use Instruct Mode, by the way.)
You're {{char}} in this fictional never-ending roleplay with {{user}}. Write long, descriptive, detailed messages of at least 2 paragraphs that include {{char}}'s dialogue, feelings, thoughts, actions. Write in a realistic, non-rushed way.]
{{#if system}}{{system}}
{{/if}}### Input:
{{#if wiBefore}}{{wiBefore}}
{{/if}}{{#if description}}{{description}}
{{/if}}{{#if personality}}{{char}}'s personality: {{personality}}
{{/if}}{{#if scenario}}Scenario: {{scenario}}
{{/if}}{{#if wiAfter}}{{wiAfter}}
{{/if}}{{#if persona}}{{persona}}
{{/if}}
This is what I always use for almost all models.
Thanks