is it Model broken?

#1
by Milor123 - opened

It print the jump lines as \n{1} where {1} could be anything number, also writes \n\t* and never stop of write, I am using mistral template in LMstudio, I've tested it with 4096 n_ctx and with 2048, but i get the same result.

Need it someone STOP word?

Temperature 0.5 and other parameters as default
image.png

Temperature 0.6 and other parameters as default
image.png

again with 0.7

image.png

Anima org

Hey! That for reaching out, this model definitely has some issues. It's more of an experiment than something that is fully tested and sturdy. It was trained on a very, very long form multi-turn dataset that went through problem solving strategies really intricately. As a by product it has verbose outputs and sometimes erratic moments. I am currently working on a new revamped version of ANIMA that is based on the Nexus models and Internal Knowledge Map dataset, so it's unlikely that I'll get around to fixing or updating this one, sorry. There are a few other GGUF's of the ANIMA model that are more stable and reliable.

Here's a little more info on the reason for why it's broken from the Model Card:

(Very longform outputs, could be broken. Still fun to use and still very clever) This model version is the 9th sequential fine-tuning of the original ANIMA-Phi-Neptune-Mistral-7B and has gone through an additional 20 hours of training using a custom dataset that combined Nectar along with my Biomimicry/STEM data in a unique interwoven way. It has A LOT of 'active' knowledge and ability to form new and novel relationships for problem-solving. It also likes to be VERY verbose, to the point of non-stop talking at points, since its' last training was on longform multi-turn problem solving design strategies in the form of conversational discourse (my custom Biomimicry/Nectar dataset). Have fun!

Sign up or log in to comment