General discussion/presets/feedback.
@localfultonextractor Heya.
Can you please test this character card:
https://www.chub.ai/characters/lewdiculous/chiaki-77cace9b
...and tell me if you can get the model to respect this formatting:
https://freeimage.host/i/Jju5Syl
https://freeimage.host/i/Jju58v4
Screenshots were from another model. Control: Kunoichi-DPO-v2-7B.
Formatting just isn't working consistently for me.
This one of the cards I use to test ability to follow formatting, and it's pretty hot as well, so there's a bit of motivation for ya.
Alpaca format.
--contextsize 8192
If you managed to, share your preset files (catbox.moe works) so I can match them.
Alright. Tested it on Mistral, Alpaca-Roleplay and ChatML default presets in SillyTavern and they all work. Your formatting problems aren't because of the chat template but because of the samplers. Mistral 0.2 and Eros like slightly lower temps. V3-0.2 follows the format perfectly with 0.95 temp, 0.1 min p and 1.05 rep pen. Include names = off
@Lewdiculous
Let me know what you think of this version. Feedback is very appreciated.
Let me know what you think of this version. Feedback is very appreciated.
"No, I won't! I'll keep the feedback all to myself! Reee...!" XD
I'm already keeping the feedback sticker on:
Aight.
I'll give your sampler parameters a try later on then. Thanks for participating actively in testing.
Thank you, @localfultonextractor, I'm impressed with how much relaxing the samplers actually did, temperature really is crucial. I am having much better results currently with the parameters you recommended. It's almost night and day, I'm used to not considering it too much as I have temperature last, but alas, I was wrong. I'll need to test some cards with more complex/static relevant formatting -- a certain Correction Shool and a certain Mom (...) RPG -- but it already looks promising.
I uploaded presets to the repo:
https://huggingface.co/Lewdiculous/Erosumika-7B-v3-0.2-GGUF-IQ-Imatrix/tree/main/sillytavern-presets-lewdicu-3.0.2-mistral-0.2
If anything seems out of place on them let me know. I'll need to take a more careful approach for 0.2 now.
A less aggressive temperature seems to really be the pattern to go for initially even with other models that share the Mistral 0.2 base now. Good stuff, I can already see improvements not just from little Sumika here, but looks like we'll be eating good in the near future. Keep up the good work.
I'll need to do a real "SumikaV3 vs InfinityRP vs Kunoichi-DPO" soon enough to attest a winner.
Also, just to add a little bit of that fire of competitiveness here, your "opponent" at the moment is Nitral's Eris_PrimeV4-Vision-32k-7B, Eris is also a pretty smart girl, more stable than ever now, also graced by the lucky stroke of Mistral 0.2, and is a good multimodal for roleplay #vision tasks, ya'know, to comment on our degenerate images, as you naturally do...
Also, just to add a little bit of that fire of competitiveness here, your "opponent" at the moment is Nitral's Eris_PrimeV4-Vision-32k-7B, Eris is also a pretty smart girl, more stable than ever now, also graced by the lucky stroke of Mistral 0.2, and is a good multimodal for roleplay #vision tasks, ya'know, to comment on our degenerate images, as you naturally do...
... I will look into adding LLaVA support for the next model. Must keep the torch ablaze.
It's always a matter of what has to give, if you need to sacrifice quality for it, generally I'd stick to focusing on text-gen and mastering that, I don't think there's a huge demand for llava support just yet from a multimodal point of view, from what I feel from users at least. Either that or they are a silent majority or not interested enough.
Oh no! What happened to our Erosumika creator! He is deleted and missing!
Unfortunately their Reddit account has also been deleted, hope they are okay :'(
Bummer. Ero was my favorite. But I moved to Noromaid for context size. The new version seems to be very spatially forgetful.