CultriX/Qwen2.5-14B-Hyperionv6

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 djuna/Q2.5-Veltha-14B-0.5 as a base.

Models Merged

The following models were included in the merge:

• hotmailuser/QwenSlerp2-14B
• sometimesanotion/Lamarck-14B-v0.6
• CultriX/SeQwence-14B-EvolMerge
• qingy2024/Fusion4-14B-Instruct
• CultriX/Qwen2.5-14B-Hyperionv3
• CultriX/Qwen2.5-14B-Brocav7
• CultriX/Qwen2.5-14B-Brocav3

Configuration

The following YAML configuration was used to produce this model:

merge_method: dare_ties
base_model: djuna/Q2.5-Veltha-14B-0.5
output_dtype: bfloat16
data_type: bfloat16
parameters:
  epsilon: 0.0085  # Balanced between precision and flexibility.
  lambda: 2.3  # Adjusted to emphasize impactful parameters without overfitting.
  normalize: true  # Ensures parameter normalization for stable integration.
  rescale: true  # Dynamically rescales parameters for optimal alignment.
  int8_mask: true  # Enables memory-efficient fine-tuning when applicable.

adaptive_merge_parameters:
  task_weights:  # Combined priority based on all configurations.
    tinyArc: 1.7
    tinyHellaswag: 1.8
    tinyMMLU: 2.0
    tinyTruthfulQA: 2.8
    tinyTruthfulQA_mc1: 2.4
    tinyWinogrande: 2.1
    IFEval: 3.2
    BBH: 2.9
    MATH: 3.4
    GPQA: 2.6
    MUSR: 2.7
    MMLU-PRO: 2.5
  smoothing_factor: 0.025  # Precise parameter blending.

gradient_clipping:  # Hybrid gradient clipping strategy.
  djuna/Q2.5-Veltha-14B-0.5: 0.89
  CultriX/Qwen2.5-14B-Brocav3: 0.88
  CultriX/Qwen2.5-14B-Hyperionv3: 0.87
  CultriX/Qwen2.5-14B-Wernickev3: 0.88
  hotmailuser/QwenSlerp2-14B: 0.90
  allknowingroger/QwenSlerp6-14B: 0.86
  sometimesanotion/Lamarck-14B-v0.6: 0.88
  qingy2024/Fusion4-14B-Instruct: 0.91
  CultriX/Qwen2.5-14B-Brocav7: 0.88
  CultriX/SeQwence-14B-EvolMerge: 0.87

models:
  - model: djuna/Q2.5-Veltha-14B-0.5
    parameters:
      weight: 0.28  # Backbone with strong reasoning capabilities.
      density: 0.78
  - model: CultriX/Qwen2.5-14B-Brocav3
    parameters:
      weight: 0.25  # High-performance reasoning and multitask contributions.
      density: 0.76  
  - model: CultriX/Qwen2.5-14B-Hyperionv3
    parameters:
      weight: 0.18  # Balanced generalist for broad coverage.
      density: 0.75
  - model: hotmailuser/QwenSlerp2-14B
    parameters:
      weight: 0.13  # Specialist in instruction-following and QA.
      density: 0.72
  - model: sometimesanotion/Lamarck-14B-v0.6
    parameters:
      weight: 0.10  # Multi-step reasoning and task-specific expert.
      density: 0.65
  - model: qingy2024/Fusion4-14B-Instruct
    parameters:
      weight: 0.08  # Specialist in mathematical reasoning.
      density: 0.78
  - model: CultriX/Qwen2.5-14B-Brocav7
    parameters:
      weight: 0.08  # Focus on specific reasoning tasks.
      density: 0.77
  - model: CultriX/SeQwence-14B-EvolMerge
    parameters:
      weight: 0.07  # Generalist for multitask integration.
      density: 0.68