CultriX/Qwen2.5-14B-Emergedv3

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:

• sometimesanotion/Qwen2.5-14B-Vimarckoso
• CultriX/SeQwence-14B-EvolMerge
• CultriX/Qwen2.5-14B-SLERPv7
• qingy2019/Qwen2.5-Math-14B-Instruct
• allknowingroger/QwenSlerp6-14B
• CultriX/Qwen2.5-14B-Wernicke
• VAGOsolutions/SauerkrautLM-v2-14b-DPO

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: VAGOsolutions/SauerkrautLM-v2-14b-DPO
    parameters:
      weight: 0.20  # Strong IFEval and factual reasoning baseline
      density: 0.6
  - model: allknowingroger/QwenSlerp6-14B
    parameters:
      weight: 0.20  # Balanced reasoning across multiple benchmarks
      density: 0.6
  - model: CultriX/SeQwence-14B-EvolMerge
    parameters:
      weight: 0.15  # Generalist model for BBH and MUSR
      density: 0.5
  - model: CultriX/Qwen2.5-14B-Wernicke
    parameters:
      weight: 0.15  # QA leader for GPQA and MUSR
      density: 0.6  # Increase density to preserve more QA-specific parameters
  - model: qingy2019/Qwen2.5-Math-14B-Instruct
    parameters:
      weight: 0.15  # Specialist for MATH and advanced reasoning
      density: 0.6
  - model: sometimesanotion/Qwen2.5-14B-Vimarckoso
    parameters:
      weight: 0.10  # MUSR leader for nuanced multi-step reasoning
      density: 0.5
  - model: CultriX/Qwen2.5-14B-SLERPv7
    parameters:
      weight: 0.05  # Contextual reasoning support for BBH and tiny benchmarks
      density: 0.5
base_model: CultriX/SeQwence-14Bv1
merge_method: dare_ties
parameters:
  normalize: true
  int8_mask: true
dtype: bfloat16
adaptive_merge_parameters:
  task_weights:
    IFEval: 1.3        # Enhanced instruction-following and factual tasks
    BBH: 1.3           # Strengthened complex reasoning capabilities
    MATH_Lvl_5: 1.4    # Prioritize advanced mathematical tasks
    GPQA: 1.4          # Boost graduate-level knowledge capabilities
    MuSR: 1.3          # Strengthen multi-step reasoning on complex tasks
    MMLU_PRO: 1.2      # Ensure broad domain understanding
  smoothing_factor: 0.15  # Sharper blending for reasoning and factual tasks
gradient_clipping: 0.9   # Tighter control for precise parameter scaling