|
--- |
|
base_model: |
|
- CultriX/Qwen2.5-14B-Wernickev3 |
|
- CultriX/Qwen2.5-14B-Emergedv3 |
|
- qingy2019/Qwen2.5-Math-14B-Instruct |
|
- CultriX/Qwen2.5-14B-FinalMerge |
|
- CultriX/SeQwence-14Bv1 |
|
library_name: transformers |
|
tags: |
|
- mergekit |
|
- merge |
|
|
|
--- |
|
# merge |
|
|
|
This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit). |
|
|
|
## Merge Details |
|
### Merge Method |
|
|
|
This model was merged using the [DARE](https://arxiv.org/abs/2311.03099) [TIES](https://arxiv.org/abs/2306.01708) merge method using [CultriX/SeQwence-14Bv1](https://huggingface.co/CultriX/SeQwence-14Bv1) as a base. |
|
|
|
### Models Merged |
|
|
|
The following models were included in the merge: |
|
* [CultriX/Qwen2.5-14B-Wernickev3](https://huggingface.co/CultriX/Qwen2.5-14B-Wernickev3) |
|
* [CultriX/Qwen2.5-14B-Emergedv3](https://huggingface.co/CultriX/Qwen2.5-14B-Emergedv3) |
|
* [qingy2019/Qwen2.5-Math-14B-Instruct](https://huggingface.co/qingy2019/Qwen2.5-Math-14B-Instruct) |
|
* [CultriX/Qwen2.5-14B-FinalMerge](https://huggingface.co/CultriX/Qwen2.5-14B-FinalMerge) |
|
|
|
### Configuration |
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
```yaml |
|
models: |
|
- model: CultriX/Qwen2.5-14B-Wernickev3 |
|
parameters: |
|
weight: 0.38 # Slight reduction to balance with FinalMerge's generalist capabilities. |
|
density: 0.65 # Retain significant parameters for stability and strong task performance. |
|
- model: CultriX/Qwen2.5-14B-FinalMerge |
|
parameters: |
|
weight: 0.32 # Slight increase to ensure its generalist capabilities are fully utilized. |
|
density: 0.60 # Balanced density for comprehensive task coverage. |
|
- model: CultriX/Qwen2.5-14B-Emergedv3 |
|
parameters: |
|
weight: 0.20 # Retains focused contribution to specific task optimizations. |
|
density: 0.55 # Moderate density ensures efficient parameter usage. |
|
- model: qingy2019/Qwen2.5-Math-14B-Instruct |
|
parameters: |
|
weight: 0.10 # Consistent with its specialist focus, balancing lower weight with higher density. |
|
density: 0.70 # High density ensures retention of advanced reasoning and MATH-related parameters. |
|
|
|
merge_method: dare_ties |
|
base_model: CultriX/SeQwence-14Bv1 |
|
parameters: |
|
normalize: true # Ensures all models are scaled to compatible parameter ranges. |
|
int8_mask: true # Optimizes memory and computational efficiency without accuracy loss. |
|
dtype: bfloat16 # Provides better memory efficiency and numerical stability. |
|
|
|
adaptive_merge_parameters: |
|
task_weights: |
|
tinyArc: 1.3 # Slight reduction to balance with generalist contributions. |
|
tinyHellaswag: 1.3 # Maintains strong performance in contextual reasoning. |
|
tinyMMLU: 1.2 # Balanced focus for domain-specific knowledge. |
|
tinyTruthfulQA: 1.2 # Adjusted to ensure fair contribution without over-prioritization. |
|
tinyTruthfulQA_mc1: 1.1 # Maintains a moderate priority to balance with other tiny benchmarks. |
|
tinyWinogrande: 1.2 # Strong contextual reasoning support from generalist models. |
|
IFEval: 1.5 # High weight for general instruction-following capabilities. |
|
BBH: 1.5 # Prioritizes complex reasoning and multi-step problem-solving tasks. |
|
MATH: 1.55 # Slight reduction to balance MATH with other advanced reasoning benchmarks. |
|
GPQA: 1.4 # Balanced to reflect contributions from both generalist and specialist models. |
|
MUSR: 1.4 # Increased slightly to strengthen multi-step reasoning. |
|
MMLU-PRO: 1.3 # Maintains general task performance across multitask domain knowledge. |
|
smoothing_factor: 0.18 # Slightly increased for smoother blending across task boundaries. |
|
gradient_clipping: 0.88 # Tightened slightly for stability, preventing parameter over-contribution. |
|
|
|
``` |
|
|
