[FeedBack]
here..
The model feels like the previous version, but with a huge context and slightly more logical (in-depth) answers. Good :)
I noticed that Ttimofeyka also created Llama-3-15B-512k-Instruct . Are you interested in doing it?
(Sorry for my bad eng)
I noticed that Ttimofeyka also created Llama-3-15B-512k-Instruct . Are you interested in doing it?
(Sorry for my bad eng)
Hey, yeah, I know. I discussed it with them, but it would just degrade the quality because the datasheet I used to train it was a maximum of 30k context.
The model still has huge issues. It starts out strong, but quickly dissolves into rambling incoherently and often exhausting the max token output which I've set to 1536. Setting Top K to 20 mitigates it a bit, but it's still a far cry from good.
This is the tail end of a typical reply. The start is usually quite good but then it turns into a mess like this:
yea i've noticed it. it really likes shorter responses try using the recommended samplers.
I did use the recommended samplers for this. Aside from XTC, which is only in the staging branch as of now. And with good models the sampler settings are flexible enough to allow for variety. Slightly wrong sampler settings shouldn't break the model.
I did use the recommended samplers for this. Aside from XTC, which is only in the staging branch as of now. And with good models the sampler settings are flexible enough to allow for variety. Slightly wrong sampler settings shouldn't break the model.
This model is based on 15B 64k - duplicating llama layers and using "hacks" to reduce model perplexity. This can lead to problems with samplers as well (which I also did see). This is mostly due to samplers related to rep pen. Even I encountered this when I tried to upload my old story (but the new story is successful). Try to find the best value for yourself - bad character cards (SillyTavern) using bad formatting can also affect this.
My system prompt: "You are an expert actor that can fully immerse yourself into any role given. You do not break character for any reason, even if someone tries addressing you as an AI or language model. Currently your role is {{char}}, which is described in detail below. As {{char}}, shortly continue the exchange with {{user}}. Talk as others NPC except {{user}} when needed."
And this is my samplers.
Thanks, I'll give those samplers a try, but I feel like the model is too finnicky for my taste. I'd love to have a L3 model with a higher context than 8k, but it seems like that's a tall order to accomplish.
Thanks, I'll give those samplers a try, but I feel like the model is too finnicky for my taste. I'd love to have a L3 model with a higher context than 8k, but it seems like that's a tall order to accomplish.
check the 2.5 version.
it's better.
And this is my samplers.
This is my setting interface. How could i get that settings interface like you?
And this is my samplers.
This is my setting interface. How could i get that settings interface like you?
Lol. Maybe you are using old version of SillyTavern?
And this is my samplers.
This is my setting interface. How could i get that settings interface like you?
Lol. Maybe you are using old version of SillyTavern?
i think that i using new version, is that right?
Right... I don't know then, lol.
i know.
And this is my samplers.
This is my setting interface. How could i get that settings interface like you?
Lol. Maybe you are using old version of SillyTavern?
i think that i using new version, is that right?
you're using chat completition. use text completition.