Update
Seemed much more verbose to me now, though I only tested a few turns. Started playing itself, making its own choices (all in one turn, 2k tokens output). Didn't look for coherence this time, but... the vibe seemed a bit off for me. Strange word choices, etc. 👀
Is this the new Ministral 14B model? That could be it... I tried it, and it sucked, also tried the 8B and it sucked. I'm no expert in LLM's, but... IME, If the base-instruct model sucks, the fine-tunes are going to be the same.
Is this the new Ministral 14B model? That could be it... I tried it, and it sucked, also tried the 8B and it sucked. I'm no expert in LLM's, but... IME, If the base-instruct model sucks, the fine-tunes are going to be the same.
It is my understanding that this is indeed the new Ministral 14B model. What about it sucked? I have no experience with those, yet...
Ime, it is creative, but makes a lot of mistakes and doesn't flow naturally, compared to Nemo. I tried the base and the instruct version of 14B.
Ime, it is creative, but makes a lot of mistakes and doesn't flow naturally, compared to Nemo. I tried the base and the instruct version of 14B.
Exactly how I felt during my little time with "Brother Dusk"! Thanks 👍
Have you guys tried it with latest llama.cpp? They merged some fixes for these latest Mistral AI models. While the bigger motivation was the Devstral 2 model, I believe it also improved the Ministral models including its finetunes. More fixes may come later still. I think it's fair to give all of these first finetunes a second chance. I already tried the Brother-Dusk-14B-v1b-GGUF again and it was pretty good.
Have you guys tried it with latest llama.cpp? They merged some fixes for these latest Mistral AI models. While the bigger motivation was the Devstral 2 model, I believe it also improved the Ministral models including its finetunes. More fixes may come later still. I think it's fair to give all of these first finetunes a second chance. I already tried the Brother-Dusk-14B-v1b-GGUF again and it was pretty good.
Thanks, will try that out!