metadata

base_model:
  - CultriX/SeQwence-14Bv1
  - CultriX/Qwen2.5-14B-Broca
  - CultriX/Qwen2.5-14B-Wernickev3
  - CultriX/Qwen2.5-14B-FinalMerge
  - sthenno-com/miscii-14b-1225
  - djuna/Q2.5-Veltha-14B
library_name: transformers
tags:
  - mergekit
  - merge

merge

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

Merge Details

Merge Method

This model was merged using the della_linear merge method using djuna/Q2.5-Veltha-14B as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

name: Merged-14B-Ultimate
merge_method: della_linear
base_model: djuna/Q2.5-Veltha-14B

dtype: bfloat16

parameters:
  epsilon: 0.01       # Fine-grained parameter scaling for stable merges
  lambda: 1.5         # Emphasizes each model’s unique parameters
  normalize: true     # Normalizes merges across different scale factors

models:
  # 1) Strong average + BBH + conversation
  - model: sthenno-com/miscii-14b-1225
    parameters:
      weight: 0.25
      density: 0.70

  # 2) CultriX “FinalMerge” synergy
  - model: CultriX/Qwen2.5-14B-FinalMerge
    parameters:
      weight: 0.15
      density: 0.65

  # 3) CultriX “Wernickev3”—balanced
  - model: CultriX/Qwen2.5-14B-Wernickev3
    parameters:
      weight: 0.15
      density: 0.65

  # 4) CultriX “Broca”—logic & QA
  - model: CultriX/Qwen2.5-14B-Broca
    parameters:
      weight: 0.10
      density: 0.65

  # 5) CultriX “SeQwence-14Bv1”—general coverage
  - model: CultriX/SeQwence-14Bv1
    parameters:
      weight: 0.10
      density: 0.65

adaptive_merge_parameters:
  # Weighted emphasis on sub-benchmarks
  task_weights:
    IFEval: 1.9
    BBH: 1.8
    MATH: 1.8
    GPQA: 1.7
    MUSR: 1.7
    MMLU-PRO: 1.7
  smoothing_factor: 0.1

gradient_clipping: 1.0   # Prevents over-contribution from any one model