my five cents on Stheno 3.4
repeats itself quite often with or without DRY Sampler 0.8/1.75/2/0.
the other values i'll took from the model card and were used as described.
significant deterioration from stheno 3.2 and higher in my opinion the NEO Variant is a bit better from it's merge used it before Nitamer v1.
Thank you very much for your time and effort
PS: Nitamer v1 is in my opinion the better choice for me
repeats itself quite often with or without DRY Sampler 0.8/1.75/2/0.
Didn't repeat itself for me in my experience, are you using Koboldcpp by any chance?. I had to switch recently to Text Generation Webui because of determinism with every model using koboldcpp, something just broke with one of the updates, at least for me, could be context shifting potentially that is the issue, was repeating itself even with DRY enabled. It's funny because like a year ago almost I had to switch from text gen to koboldcpp because of a determinism bug, so it's just back and forth I guess lol. Otherwise the coherency is definitely worse over Stheno 3.2
repeats itself quite often with or without DRY Sampler 0.8/1.75/2/0.
Didn't repeat itself for me in my experience, are you using Koboldcpp by any chance?. I had to switch recently to Text Generation Webui because of determinism with every model using koboldcpp, something just broke with one of the updates, at least for me, could be context shifting potentially that is the issue, was repeating itself even with DRY enabled. Otherwise the coherency is definitely worse over Stheno 3.2
yes i use koboldcpp and it does it's job nitama v1 and older stheno 3.1 runs fine 4 me with 12288 context size everything is fine so far. kobold 1.73 released yesterday so far it's stable last versions had no problems so. nitama v1 never gave me such repeating phrases over and over again maybe the crossover merge to llama 3.1 was the clincher.
Need to clarify, latest koboldcpp has broken DRY samplers and devs are working to fix the issue, so it's not a model problem.
This is the statement on Koboldcpp release page right now for newest version.
NOTICE: DRY is completely broken in this version. It will also cause issues in ST. A fix is currently being developed - for now do not use DRY or switch to 1.72
I feel like DRY was not working in 1.72 either, because I was using it and feeling almost no difference compared to text gen webui. Guess something was messed up after all.
Need to clarify, latest koboldcpp has broken DRY samplers and devs are working to fix the issue, so it's not a model problem.
yeah but u can run it with or
w i t h o u t
DRY sampling
Same same..... Probs are constant, at least for me,
but no offense
This is the statement on Koboldcpp release page right now for newest version.
NOTICE: DRY is completely broken in this version. It will also cause issues in ST. A fix is currently being developed - for now do not use DRY or switch to 1.72
I feel like DRY was not working in 1.72 either, because I was using it and feeling almost no difference compared to text gen webui. Guess something was messed up after all.
u can pull the plug on that dry stuff in sillytav
idk how that spill over or not
that was my way to test it out and gave me the same weird output as i wrote in the first post.
hm... weird, i've tried LM Studio and the output changes in a pretty decent kinda way.
no weird repetitions at all ....
it seems like koboldcpp was the problem after all
thx everyone from above