|
--- |
|
license: other |
|
license_name: mrl |
|
language: |
|
- en |
|
tags: |
|
- chat |
|
pipeline_tag: text-generation |
|
library_name: transformers |
|
base_model: |
|
- MarsupialAI/Monstral-123B-v2 |
|
base_model_relation: quantized |
|
quantized_by: BigHuggyD |
|
--- |
|
# Monstral 123B v2 |
|
A Mistral-Large merge |
|
 |
|
|
|
This model is a hybrid merge of Behemoth 1.2, Tess, and Magnum V4. The intention was to do a three-way slerp merge, which is technically |
|
not possible. To simulate the effeect of a menage-a-slerp, I slerped B1.2 with tess, then separately did B1.2 with magnum. I then did a |
|
model stock merge of those two slerps using B1.2 as the base. Somehow, it worked out spectacularly well. Sometimes dumb ideas pay off. |
|
|
|
Mergefuel: |
|
- TheDrummer/Behemoth-123B-v1.2 |
|
- anthracite-org/magnum-v4-123b |
|
- migtissera/Tess-3-Mistral-Large-2-123B |
|
|
|
See recipe.txt for full details. |
|
|
|
Improvements over Monstral v1: Drummer's 1.2 tune of behemoth is a marked improvement over the original, and the addition ot tess to the |
|
mix really makes the creativity pop. I seem to have dialed out the rapey magnum influence, without stripping it of the ability to get mean |
|
and/or dirty when the situation actually calls for it. The RP output of this model shows a lot more flowery and "literary" description of |
|
scenes and activities. It's more colorful and vibrant. Repitition is dramatically reduced, as is slop (though to a lesser extent). The |
|
annoying tendency to double-describe things with "it was X, almost Y" is virtually gone. Do you like a slow-burn story that builds over |
|
time? Well good fucking news, because v2 excels at that. |
|
|
|
The only complaint I've received is occasional user impersonation with certain cards. I've not seen this myself on any of my cards, so I |
|
have to assume it's down to the specific formatting on specific cards. I don't want to say it's a skill issue, but... |
|
|
|
This model is uncensored and perfectly capable of generating objectionable material. I have not observed it injecting NSFW content into |
|
SFW scenarios, but no guarentees can be made. As with any LLM, no factual claims made by the model should be taken at face value. You |
|
know that boilerplate safety disclaimer that most professional models have? Assume this has it too. This model is for entertainment |
|
purposes only. |
|
|
|
GGUFs: https://huggingface.co/MarsupialAI/Monstral-123B-v2_GGUF |
|
|
|
|
|
# Prompt Format |
|
Metharme seems to work flawlessly. In theory, mistral V3 or possibly even chatml should work to some extent, but meth was providing such |
|
high quality output that I couldn't even be bothered to test the others. Just do meth, kids. |
