A short and simple review from an average observer.
I'm not the type who participates so directly in the LLM Models community, but that's been changing in recent years, thankfully :)
And with that introduction about myself I'll begin:
I wasn't expecting such promising results, how can I explain... generally the cards that many people write or download are of (optimistically) average quality, resulting in... bad and poor results in the worst cases, especially if we're talking about two characters on the same card.
In L3-8B-Stheno-v3.2 the results were very average or Bad when there were two characters in the context or card, but L3-8B-Stheno-v3.3-32K managed to produce an effect close to what I only see with MOEs, of course there are errors but I see a future in this method. In the old AID times it was normal for me to do RPG adventures, this has become more difficult in LLM, but I think it will be possible soon. @Sao10K I hope my feedback can help you in your future improvements.
I'm not the type who participates so directly in the LLM Models community, but that's been changing in recent years, thankfully :)
And with that introduction about myself I'll begin:
I wasn't expecting such promising results, how can I explain... generally the cards that many people write or download are of (optimistically) average quality, resulting in... bad and poor results in the worst cases, especially if we're talking about two characters on the same card.In L3-8B-Stheno-v3.2 the results were very average or Bad when there were two characters in the context or card, but L3-8B-Stheno-v3.3-32K managed to produce an effect close to what I only see with MOEs, of course there are errors but I see a future in this method. In the old AID times it was normal for me to do RPG adventures, this has become more difficult in LLM, but I think it will be possible soon. @Sao10K I hope my feedback can help you in your future improvements.
What settings do you use? 3.2, for some reason, out performs 3.3 in all measures. It just seems to be more consistent and in-line on 3.2 while 3.3 seems to bounce around.
What settings do you use? 3.2, for some reason, out performs 3.3 in all measures. It just seems to be more consistent and in-line on 3.2 while 3.3 seems to bounce around.
I used the recommendations from the comments on Lewdiculous' Imatrix GGUF Community post, which works perfectly well with L3-8B-Stheno-v3.1 & 2, and one thing I also experienced was some repetition at the end of sentences.
Those are the settings link:
https://huggingface.co/Virt-io/SillyTavern-Presets/tree/main/Prompts/LLAMA-3/v1.9
https://files.catbox.moe/78inw0.json