🩵 LunaMaid-12B

This is a multi-stage merge of pre-trained language models created using mergekit.

🧬 Merge Overview

LunaMaid-12B was produced through a two-stage multi-model merge using MergeKit.
Each stage fuses models with complementary linguistic and stylistic traits to create a cohesive, emotionally nuanced personality.

🩵 Stage 1 — Slerp Merge (Intermediate Model `First`)

Base Model: Vortex5/Vermilion-Sage-12B
Merged With: yamatazen/NeonMaid-12B-v2
Method: Spherical Linear Interpolation (Slerp)

Stage 1 Configuration

name: First
base_model: Vortex5/Vermilion-Sage-12B
models:
- model: yamatazen/NeonMaid-12B-v2
merge_method: slerp
dtype: bfloat16
parameters:
normalize: true
t: [0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.6, 0.5, 0.6, 0.6]

🌑 Merge Method — Karcher Mean Merge (Final Model)

Base Model: Intermediate output from Stage 1 ./intermediates/First
Merged With: Vortex5/Moonlit-Shadow-12B
Method: Karcher Mean (Riemannian Barycenter)

Stage 2 Configuration

dtype: bfloat16
merge_method: karcher
modules:
  default:
    slices:
      - sources:
          - layer_range: [0, 40]
            model: ./intermediates/First
          - layer_range: [0, 40]
            model: Vortex5/Moonlit-Shadow-12B
parameters:
  max_iter: 9999
  tol: 1e-9