merge
This is a merge of pre-trained language models created using mergekit.
Merge Details
Merge Method
This model was merged using the della_linear merge method using CultriX/Qwen2.5-14B-Wernickev3 as a base.
Models Merged
The following models were included in the merge:
- djuna/Q2.5-Veltha-14B-0.5
- CultriX/Qwenfinity-2.5-14B
- sometimesanotion/Lamarck-14B-v0.6
- CultriX/Qwen2.5-14B-Emerged
- CultriX/Qwen2.5-14B-Broca
- qingy2024/Fusion4-14B-Instruct
- CultriX/SeQwence-14B-EvolMerge
- allknowingroger/QwenSlerp5-14B
Configuration
The following YAML configuration was used to produce this model:
merge_method: della_linear
base_model: CultriX/Qwen2.5-14B-Wernickev3
dtype: bfloat16
out_dtype: bfloat16
parameters:
epsilon: 0.009 # Further reduced for ultra-fine parameter scaling.
lambda: 1.6 # Increased to emphasize significant model contributions.
normalize: true # Balances the parameter integration for stability.
rescale: true # Enabled to align parameter scales across models.
int8_mask: false # Disabled to allow full-precision computations for enhanced accuracy.
density: 0.90 # Balanced density for optimal generalization and performance.
adaptive_merge_parameters:
task_weights:
tinyArc: 1.6 # Prioritizes logical reasoning improvements.
tinyHellaswag: 1.5 # Strengthened contextual understanding and consistency.
tinyMMLU: 1.8 # Enhanced domain knowledge for multitask benchmarks.
tinyTruthfulQA: 1.9 # Maximized for accurate factual reasoning and QA.
tinyTruthfulQA_mc1: 1.75 # Increased focus for multiple-choice reasoning.
tinyWinogrande: 1.75 # Advanced reasoning and contextual prediction improvement.
IFEval: 2.15 # Enhanced instruction-following tasks boosted by multitask contributors.
BBH: 1.95 # Further improved for complex reasoning tasks.
MATH: 2.45 # Highest priority, focusing on mathematical excellence.
GPQA: 2.1 # Boosted graduate-level QA capabilities.
MUSR: 1.9 # Nuanced multi-step reasoning strengthened further.
MMLU-PRO: 1.9 # Maximized domain multitask performance.
smoothing_factor: 0.035 # Further reduced for precise task-specific blending.
gradient_clipping:
CultriX/Qwen2.5-14B-Wernickev3: 0.88 # Increased for enhanced stability.
CultriX/Qwenfinity-2.5-14B: 0.85 # Adjusted for consistent multitask integration.
djuna/Q2.5-Veltha-14B-0.5: 0.91 # Maintained advanced reasoning contributions.
CultriX/SeQwence-14B-EvolMerge: 0.88 # Generalist multitask support remains stable.
qingy2024/Fusion4-14B-Instruct: 0.93 # Mathematically focused tasks maximized.
CultriX/Qwen2.5-14B-Emerged: 0.88 # Increased for logical reasoning enhancements.
sometimesanotion/Lamarck-14B-v0.6: 0.89 # Balanced multi-step reasoning contributions.
allknowingroger/QwenSlerp5-14B: 0.87 # Contextual and logical reasoning integration refined.
models:
- model: CultriX/Qwen2.5-14B-Wernickev3
parameters:
weight: 0.30 # Core backbone for multitask reasoning.
density: 0.75 # Further increased to preserve critical reasoning parameters.
- model: CultriX/Qwenfinity-2.5-14B
parameters:
weight: 0.25 # Comprehensive multitask performer.
density: 0.65
- model: djuna/Q2.5-Veltha-14B-0.5
parameters:
weight: 0.25 # Advanced reasoning support for GPQA and MUSR.
density: 0.74
- model: CultriX/SeQwence-14B-EvolMerge
parameters:
weight: 0.20 # Enhanced contributions to BBH and MUSR.
density: 0.55
- model: qingy2024/Fusion4-14B-Instruct
parameters:
weight: 0.19 # Mathematical reasoning priority.
density: 0.77
- model: CultriX/Qwen2.5-14B-Emerged
parameters:
weight: 0.21 # Maintains overall task performance with balanced strengths.
density: 0.72 # Increased for better integration.
- model: CultriX/Qwen2.5-14B-Broca
parameters:
weight: 0.16 # Logical reasoning and factual QA enhancements.
density: 0.68 # Increased to better support Broca's specialized tasks.
- model: sometimesanotion/Lamarck-14B-v0.6
parameters:
weight: 0.15 # Multi-step reasoning tasks contributor.
density: 0.63 # Slight increase for better integration.
- model: allknowingroger/QwenSlerp5-14B
parameters:
weight: 0.16 # Contextual reasoning improvements.
density: 0.64 # Increased for enhanced performance.