DavidAU/MN-DARKEST-Universe-Infinity-Grand-Horror-Instruct-29B

MN-DARKEST-Universe-Infinity-Grand-Horror-Instruct-29B

Mega fine tune of DARKEST UNIVERSE 29B; with Grand Horror dataset via Unsloth.

Model contains 102 layers and 944 tensors for superior generation ablitilies. It is custom built Mistral Nemo structure (Three custom 18.5B models, merged plus Brainstorm 40x by DavidAU) that works with llamacpp (gguf), and all other standard transformer type arches / quantizations.

Note that Token/Second speed will be lower than other 30B dense models, as this model is over 2 times as dense. This model is about quality over speed.

Flash attention can be used to speed up generation.

You can also use lower quants (Q4ks regular or higher / IQ3_M imatrix or higher) and get excellent quality too.

SUGGESTED SETTINGS:

• Temp .25 to 2.5 ; with 1.2 to 2 recommended.
• Top p .95, min p .05, top k 40 to 120.
• Rep pen 1 to 1.02.
• Context of 8k to 16k recommended.
• System prompt for enhanced performance. See the "older" GGUF repo below for creative system prompt.

Context 128k-256k, with up to 1 million possible.

This is the "instruct" version.

This version is completely stable (Class 1).

Fine tuning done via Unsloth.

Model also supports system prompt which affects both thinking and output generation.

For Claude Deep Reasoning ("thinking") version go here:

https://huggingface.co/DavidAU/MN-DARKEST-Universe-Infinity-Claude-High-Reasoning-29B

New model card coming...

NOTE:

Org model was a class 3-4 (special settings); whereas this new model is a class 1 (meaning no special settings).

See org model here for details in the meantime:

https://huggingface.co/DavidAU/MN-DARKEST-UNIVERSE-29B-GGUF