GGUF
Not-For-All-Audiences
Edit model card

tldr;

I tried merging and got VERY good results with logic & writing. Try them out!

If you do try it out, please let me know your thoughts! Just tell me which percentages you tried and how they performed. The future of Moistral is in your hands!

Moistral 11B v1, dehumidified

image/gif

Original unmerged model: https://huggingface.co/TheDrummer/Moistral-11B-v1

What is this?

GGUF merges of Moistral 11B v1 and Fimbulvetr v2.

Why merge?

The finetuned model had some problems with writing, logic, and formatting. Merging itself with the base model (Fimbul) fixed those problems while retaining a lot of its original moist.

I can confidently say that this Moistral is better than the original v1 in all aspects (moist, logic, writing, formatting).

Which one do I pick?

I'm releasing 5 versions of the merge. The % in the filename represents how much "Fimbulvetr v2" I merged into my finetuned model. The higher the percent, the less moist there is.

Assessment:

  • 50%: Barely moist. Moistral's style rarely shows up but it may have some impact to the moist moments.
  • 25%: Hints of moist. Moistral's style shows up from time to time. It cooks very slowly and the moist moments may not be cooked with the signature moist.
  • 10%: Moisty. Moistral's style is emergent. Cooking duration varies here, but there's usually some pacing involved. Moist moments have a good amount of the signature moist.
  • 5%: Quite moist. Moistral's style is apparent. Cooks quickly half the time and it definitely has the signature Moistral moist writing. Logic is still very strong at this point.
  • 2.5%: (Not Recommended) Very moist. Moistral's style is dominant but still coherent. Very picky with the prompts, it seems. Cooks all the time. Mositral problems start to emerge very noticably. Logic is faltering, with one foot out the door.

Summary

  • 50% and 25% are seemingly more Fimbulvetr than Moistral. It might not be a bad thing if you want Fimbulvetr with some extra flavor.
  • 10% and 5% if you want the MOIST. 5% if you want very moist logic and writing, 10% if you just want a lot of moist moments.
  • 2.5% is MOIST at its purest, sometimes functional state... That is, if you can get it to work properly. YMMV, broken half of the time with the wrong prompt - and when it works, it's as if a tweaked up, schizo writer from Brazz3rs wrote the story (Fun!).

Nutshell: All of them are very coherent. Lower percentages = More moist logic & writing.

Added observation: It seems like no matter how moist the situation is, the character reactions are often grounded / realistic. This applies even to 2.5%.

Note: My assessment was flawed, brief, and subjective.

"I found the perfect merge ratio!"

Great! Let me know which one and why. This WILL affect future development.

Downloads last month
794
GGUF
Model size
10.7B params
Architecture
llama
Unable to determine this model's library. Check the docs .

Collection including TheDrummer/Moistral-11B-v1.1-MERGED-PCT-GGUF