metadata

base_model:
  - CultriX/SeQwence-14B-EvolMerge
  - allknowingroger/QwenStock3-14B
  - VAGOsolutions/SauerkrautLM-v2-14b-DPO
  - allknowingroger/QwenSlerp6-14B
  - CultriX/Qwen2.5-14B-Wernicke
  - CultriX/SeQwence-14Bv1
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 DARE TIES merge method using CultriX/SeQwence-14Bv1 as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: VAGOsolutions/SauerkrautLM-v2-14b-DPO
    parameters:
      weight: 0.25    # Prioritize top IFEval
      density: 0.6     # Keep a large portion for strong factual baseline

  - model: allknowingroger/QwenSlerp6-14B
    parameters:
      weight: 0.25    # High weight for MATH and balanced reasoning
      density: 0.6     # Retain robust reasoning capabilities

  - model: CultriX/SeQwence-14B-EvolMerge
    parameters:
      weight: 0.20    # Important for best BBH and near-top MuSR
      density: 0.5     # Moderate density to ensure these strengths blend well

  - model: CultriX/Qwen2.5-14B-Wernicke
    parameters:
      weight: 0.15    # Adds top GPQA performance
      density: 0.5     # Sufficient to preserve QA strengths

  - model: allknowingroger/QwenStock3-14B
    parameters:
      weight: 0.15    # For top MMLU-PRO, enhancing domain knowledge
      density: 0.5     # Balanced integration of diverse subject expertise

base_model: CultriX/SeQwence-14Bv1
merge_method: dare_ties
parameters:
  normalize: true      # Ensures parameter scaling compatibility
  int8_mask: true      # Memory and computational efficiency
dtype: bfloat16
adaptive_merge_parameters:
  task_weights:
    IFEval: 1.2        # Emphasize instruction-following and formatting adherence
    BBH: 1.2           # Maintain strong performance in challenging reasoning tasks
    MATH_Lvl_5: 1.3    # Ensure domain expertise in competitive math problems
    GPQA: 1.3          # Leverage graduate-level knowledge capabilities
    MuSR: 1.1          # Enhance multistep reasoning on complex tasks
    MMLU_PRO: 1.2      # Ensure robust multitask domain understanding
  smoothing_factor: 0.2  # Moderate blending for stable integration
gradient_clipping: 1.0   # Prevent over-contribution from any single model