metadata

base_model:
  - qingy2024/Fusion4-14B-Instruct
  - sometimesanotion/Lamarck-14B-v0.6
  - allknowingroger/QwenSlerp6-14B
  - CultriX/SeQwence-14B-EvolMerge
  - CultriX/Qwen2.5-14B-Wernickev3
  - hotmailuser/QwenSlerp2-14B
  - CultriX/Qwen2.5-14B-Emerged
  - djuna/Q2.5-Veltha-14B-0.5
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/Qwen2.5-14B-Wernickev3 as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

merge_method: dare_ties # Changed to dare_ties
base_model: CultriX/Qwen2.5-14B-Wernickev3
dtype: bfloat16 # Use float32 for maximum precision.
out_dtype: bfloat16  # Output model also uses bfloat16 for consistency and reduced memory usage.

parameters:
  t: 0.5  # Balances interpolation between models; 0.5 gives equal weight to all contributors.
  normalize: true  # Ensures parameters are normalized to maintain stability during merging.
  rescale: true  # Aligns parameter scales across models for better integration.
  int8_mask: false  # Disable int8 masking to preserve full precision during merging.
  epsilon: 0.008  # Ultra-fine parameter scaling for precise adjustments between models.
  lambda: 1.8  # Emphasizes high-impact parameters, giving more weight to significant contributors.

adaptive_merge_parameters:
  task_weights:  # Assign weights to tasks based on their priority and impact on benchmarks.
    tinyArc: 1.6  # Logical reasoning benchmark; slightly lower priority.
    tinyHellaswag: 1.5  # Contextual reasoning benchmark with moderate priority.
    tinyMMLU: 1.8  # Multi-domain knowledge benchmark; important for multitask performance.
    tinyTruthfulQA: 1.9  # Focuses on factual reasoning and QA; high priority.
    tinyTruthfulQA_mc1: 1.75  # Multiple-choice factual reasoning; closely related to TruthfulQA.
    tinyWinogrande: 1.75  # Core reasoning benchmark; slightly lower than BBH.
    IFEval: 2.30  # Instruction-following tasks; given a high priority for practical applications.
    BBH: 2.05  # Complex reasoning benchmark; critical for logical tasks.
    MATH: 2.70  # Highest priority to emphasize mathematical reasoning excellence.
    GPQA: 2.20  # Graduate-level QA tasks; balanced priority for high-level reasoning.
    MUSR: 2.15  # Multi-step reasoning; slightly increased to strengthen reasoning performance.
    MMLU-PRO: 2.00  # Domain multitask benchmark; maintained for general multitask capability.
  smoothing_factor: 0.03  # Low smoothing for precise task-specific blending without over-generalizing.

gradient_clipping:  # Control gradient clipping for each model to stabilize training.
  CultriX/Qwen2.5-14B-Wernickev3: 0.89  # Higher value ensures stability for the base model.
  djuna/Q2.5-Veltha-14B-0.5: 0.92  # Stable setting to enhance reasoning contributions.
  CultriX/SeQwence-14B-EvolMerge: 0.87  # Moderate value for generalist multitask support.
  qingy2024/Fusion4-14B-Instruct: 0.93  # High stability to emphasize mathematical tasks.
  CultriX/Qwen2.5-14B-Emerged: 0.88  # Stable setting to maintain multitask performance.
  sometimesanotion/Lamarck-14B-v0.6: 0.89  # Stable contribution for multi-step reasoning.
  allknowingroger/QwenSlerp6-14B: 0.90  # Adjusted for stable integration of the replacement model.
  hotmailuser/QwenSlerp2-14B: 0.91  # Increased slightly for stable integration of reasoning contributions.

models:  # Define models to include in the merge, along with their weights and densities.
  - model: CultriX/Qwen2.5-14B-Wernickev3
    parameters:
      weight: 0.33  # Increased to absorb some of the weight from the removed model.
      density: 0.78  # Maintained optimal density for robust generalist performance.

  - model: djuna/Q2.5-Veltha-14B-0.5
    parameters:
      weight: 0.28  # Increased slightly to enhance reasoning benchmarks like MUSR.
      density: 0.77  # Maintained for strong nuanced reasoning.

  - model: allknowingroger/QwenSlerp6-14B  # Replacement for Qwenfinity-2.5-14B.
    parameters:
      weight: 0.15  # Matches the weight of the replaced model to preserve balance.
      density: 0.70  # Increased slightly for stronger parameter integration.

  - model: CultriX/SeQwence-14B-EvolMerge
    parameters:
      weight: 0.12  # Moderate weight for general multitask support.
      density: 0.62  # Maintained for stable contribution.

  - model: qingy2024/Fusion4-14B-Instruct
    parameters:
      weight: 0.09  # Moderate weight; focuses on mathematical reasoning tasks.
      density: 0.75  # Maintained density for stable integration.

  - model: CultriX/Qwen2.5-14B-Emerged
    parameters:
      weight: 0.08  # Balanced weight for multitask contributions.
      density: 0.69  # Maintained density for stable integration.

  - model: sometimesanotion/Lamarck-14B-v0.6
    parameters:
      weight: 0.06  # Lower weight to allow more impactful models to dominate.
      density: 0.62  # Maintained for stable multi-step reasoning contribution.

  - model: hotmailuser/QwenSlerp2-14B
    parameters:
      weight: 0.11  # Increased slightly to balance contributions.
      density: 0.66  # Maintained for stable parameter integration.