LuminumMistral-123B

Main model page: https://huggingface.co/FluffyKaeloky/Luminum-v0.1-123B

Overview

I present Luminum-123B.

This is a merge using Mistral Large as a base, and including Lumimaid-v0.2-123B and Magnum-v2-123B. I felt like Magnum was rambling too much, and Lumimaid lost slightly too much brain power, so I used Mistral Large base for a long while, but it was lacking some moist.

On a whim, I decided to merge both Lumimaid and Magnum on top of Mistral large, and while I wasn't expecting much, I've been very surprised with the results. I've found that this model keeps the brain power from Mistral base, and also inherits the lexique of Lumimaid and creative descriptions of Magnum, without rambling too much.

I've tested this model quite extensively at and above 32k with great success. It should in theory allow for the full 128k context, albeit I've only went to 40-50k max. It's become my new daily driver.

The only negative thing I could find is that it tends to generate long responses if you let it. It probably gets that from Magnum. Just don't let it grow its answer size over and over.

I recommend thoses settings:

Minp: 0.08
Rep penalty: 1.03
Rep penalty range : 4096
Smoothing factor: 0.23
No Repeat NGram Size: 2

I didn't get the chance to mess with DRY yet.

Template

All the merged models use Mistral template, this one too.

<s>[INST] {input} [/INST] {output}</s>

Quants

GGUF

Static quants IMat

EXL2

4.0bpw 5.5bpw

Merge Method

This model was merged using the della_linear merge method using mistralaiMistral-Large-Instruct-2407 as a base.

Models Merged

The following models were included in the merge: NeverSleepLumimaid-v0.2-123B anthracite-orgmagnum-v2-123b

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: anthracite-orgmagnum-v2-123b
    parameters:
      weight: 0.19
      density: 0.5
  - model: NeverSleepLumimaid-v0.2-123B
    parameters:
      weight: 0.34
      density: 0.8
merge_method: della_linear
base_model: mistralaiMistral-Large-Instruct-2407
parameters:
  epsilon: 0.05
  lambda: 1
  int8_mask: true
dtype: bfloat16