metadata

base_model:
  - allknowingroger/QwenSlerp6-14B
  - sometimesanotion/Qwen2.5-14B-Vimarckoso
  - CultriX/Qwen2.5-14B-Emergedv3
  - qingy2019/Qwen2.5-Math-14B-Instruct
  - CultriX/Qwen2.5-14B-Wernickev3
  - 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:

base_model: CultriX/SeQwence-14Bv1
merge_method: dare_ties
parameters:
  normalize: true
  int8_mask: true
dtype: bfloat16

models:
  - model: CultriX/SeQwence-14Bv1
    parameters:
      weight: 0.28      # Strong base for multitask benchmarks.
      density: 0.7      # Retains strong multitask performance.

  - model: CultriX/Qwen2.5-14B-Wernickev3
    parameters:
      weight: 0.22      # Balanced to support reasoning-heavy benchmarks like BBH.
      density: 0.65

  - model: qingy2019/Qwen2.5-Math-14B-Instruct
    parameters:
      weight: 0.22      # Optimized for MATH and BBH.
      density: 0.6

  - model: allknowingroger/QwenSlerp6-14B
    parameters:
      weight: 0.18      # Reintegration of the highest scorer for stability across benchmarks.
      density: 0.65     # Focused on its exceptional multitask and reasoning strengths.

  - model: CultriX/Qwen2.5-14B-Emergedv3
    parameters:
      weight: 0.15      # Maintains multitask stability for GPQA and MMLU-PRO.
      density: 0.6

  - model: sometimesanotion/Qwen2.5-14B-Vimarckoso
    parameters:
      weight: 0.1       # Late-layer contributor for MUSR and multi-step reasoning.
      density: 0.6

adaptive_merge_parameters:
  task_weights:
    IFEval: 1.4         # Balanced to maintain instruction-following benchmarks.
    BBH: 1.4            # Ensures strong reasoning capabilities.
    MATH: 1.5           # Prioritizes mathematical reasoning.
    GPQA: 1.5           # Balanced for factual QA.
    MUSR: 1.4           # Advanced multi-step reasoning.
    MMLU-PRO: 1.5       # Emphasized for domain-specific multitask performance.

  smoothing_factor: 0.12  # Smooth transitions between task-specific contributions.

gradient_clipping:
  CultriX/SeQwence-14Bv1: 0.8
  CultriX/Qwen2.5-14B-Wernickev3: 0.8
  qingy2019/Qwen2.5-Math-14B-Instruct: 0.85
  allknowingroger/QwenSlerp6-14B: 0.8  # Balanced for high scoring model contributions.
  CultriX/Qwen2.5-14B-Emergedv3: 0.75
  sometimesanotion/Qwen2.5-14B-Vimarckoso: 0.75

tokenizer_source: CultriX/SeQwence-14Bv1