|
--- |
|
base_model: |
|
- allknowingroger/QwenSlerp6-14B |
|
- CultriX/Qwenfinity-2.5-14B |
|
- CultriX/Qwen2.5-14B-Wernickev3 |
|
- djuna/Q2.5-Veltha-14B-0.5 |
|
- CultriX/SeQwence-14Bv1 |
|
- qingy2019/Qwen2.5-Math-14B-Instruct |
|
- sometimesanotion/Qwen2.5-14B-Vimarckoso |
|
- CultriX/Qwen2.5-14B-Broca |
|
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 della_linear merge method using [CultriX/Qwen2.5-14B-Wernickev3](https://huggingface.co/CultriX/Qwen2.5-14B-Wernickev3) as a base. |
|
|
|
### Models Merged |
|
|
|
The following models were included in the merge: |
|
* [allknowingroger/QwenSlerp6-14B](https://huggingface.co/allknowingroger/QwenSlerp6-14B) |
|
* [CultriX/Qwenfinity-2.5-14B](https://huggingface.co/CultriX/Qwenfinity-2.5-14B) |
|
* [djuna/Q2.5-Veltha-14B-0.5](https://huggingface.co/djuna/Q2.5-Veltha-14B-0.5) |
|
* [CultriX/SeQwence-14Bv1](https://huggingface.co/CultriX/SeQwence-14Bv1) |
|
* [qingy2019/Qwen2.5-Math-14B-Instruct](https://huggingface.co/qingy2019/Qwen2.5-Math-14B-Instruct) |
|
* [sometimesanotion/Qwen2.5-14B-Vimarckoso](https://huggingface.co/sometimesanotion/Qwen2.5-14B-Vimarckoso) |
|
* [CultriX/Qwen2.5-14B-Broca](https://huggingface.co/CultriX/Qwen2.5-14B-Broca) |
|
|
|
### Configuration |
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
```yaml |
|
merge_method: della_linear |
|
base_model: CultriX/Qwen2.5-14B-Wernickev3 |
|
dtype: bfloat16 |
|
parameters: |
|
epsilon: 0.01 # Reduced from 0.012 for even finer parameter scaling, enhancing precision in blending. |
|
lambda: 1.5 # Increased from 1.4 to further emphasize significant model contributions, particularly from specialized models. |
|
normalize: true # Maintains balanced parameter integration, crucial for stability across diverse benchmarks. |
|
|
|
adaptive_merge_parameters: |
|
task_weights: |
|
tinyArc: 1.65 # Increased from 1.6 to further boost logical reasoning, leveraging Qwen2.5-14B-Broca's strength. |
|
tinyHellaswag: 1.55 # Slightly increased from 1.5 to enhance contextual understanding, supported by SeQwence-14Bv1. |
|
tinyMMLU: 1.7 # Increased from 1.65 for improved domain knowledge, utilizing Qwenfinity-2.5-14B's broad capabilities. |
|
tinyTruthfulQA: 1.95 # Slightly increased from 1.9 to maximize accurate factual reasoning, with Qwenfinity-2.5-14B's contribution. |
|
tinyTruthfulQA_mc1: 1.75 # Increased from 1.7 for enhanced multiple-choice reasoning, supported by Qwen2.5-14B-Emergedv3. |
|
tinyWinogrande: 1.8 # Increased from 1.75 for advanced reasoning and contextual prediction, leveraging Qwen2.5-14B-Broca. |
|
IFEval: 2.0 # Increased from 1.9 to prioritize instruction-following, with Q2.5-Veltha-14B-0.5's strong performance. |
|
BBH: 1.75 # Slightly increased from 1.7 for complex reasoning, supported by SeQwence-14B-EvolMerge's strength. |
|
MATH: 2.2 # Increased from 2.1 to maximize mathematical reasoning, with Qwen2.5-Math-14B-Instruct's specialization. |
|
GPQA: 1.85 # Increased from 1.8 for enhanced graduate-level QA, leveraging Qwen2.5-14B-Wernicke's capabilities. |
|
MUSR: 1.95 # Increased from 1.9 for strengthened multi-step reasoning, with Qwen2.5-14B-Vimarckoso's expertise. |
|
MMLU-PRO: 1.85 # Increased from 1.8 to further boost domain multitask performance, utilizing QwenSlerp6-14B. |
|
smoothing_factor: 0.08 # Reduced from 0.1 for more precise blending, allowing distinct model strengths to be preserved. |
|
|
|
gradient_clipping: |
|
CultriX/Qwen2.5-14B-Wernickev3: 0.85 # Slightly reduced from 0.86 to allow a bit more contribution from the base model. |
|
CultriX/Qwenfinity-2.5-14B: 0.82 # Reduced from 0.83 to balance its broad multitask contribution. |
|
djuna/Q2.5-Veltha-14B-0.5: 0.92 # Slightly increased from 0.91 to allow more contribution in advanced reasoning. |
|
CultriX/Qwen2.5-14B-Broca: 0.86 # Slightly increased from 0.85 to leverage its logical reasoning strengths. |
|
qingy2019/Qwen2.5-Math-14B-Instruct: 0.94 # Increased from 0.93 to maximize its mathematical reasoning contribution. |
|
CultriX/SeQwence-14Bv1: 0.87 # Slightly reduced from 0.88 to balance its generalist multitask support. |
|
sometimesanotion/Qwen2.5-14B-Vimarckoso: 0.90 # Increased from 0.89 for enhanced multi-step reasoning. |
|
allknowingroger/QwenSlerp6-14B: 0.86 # Slightly reduced from 0.87 to refine its contextual reasoning integration. |
|
|
|
models: |
|
- model: CultriX/Qwen2.5-14B-Wernickev3 |
|
parameters: |
|
weight: 0.25 # Slightly reduced from 0.26 to balance with other models while maintaining a strong foundation. |
|
density: 0.72 # Increased from 0.7 to preserve more of its critical reasoning parameters. |
|
- model: CultriX/Qwenfinity-2.5-14B |
|
parameters: |
|
weight: 0.22 # Slightly reduced from 0.23 for a more balanced contribution across its broad capabilities. |
|
density: 0.68 # Increased from 0.65 to retain more of its multitask performance. |
|
- model: djuna/Q2.5-Veltha-14B-0.5 |
|
parameters: |
|
weight: 0.20 # Reduced from 0.22 to balance its specialized contributions with the overall blend. |
|
density: 0.75 # Increased from 0.72 to further leverage its strengths in IFEval and advanced reasoning. |
|
- model: CultriX/Qwen2.5-14B-Broca |
|
parameters: |
|
weight: 0.16 # Slightly increased from 0.15 to enhance its logical reasoning and factual QA contributions. |
|
density: 0.68 # Increased from 0.65 to preserve more of its capabilities in the tiny benchmarks. |
|
- model: qingy2019/Qwen2.5-Math-14B-Instruct |
|
parameters: |
|
weight: 0.19 # Slightly increased from 0.18 to further emphasize mathematical reasoning. |
|
density: 0.75 # Increased from 0.73 to retain more of its specialized mathematical parameters. |
|
- model: CultriX/SeQwence-14Bv1 |
|
parameters: |
|
weight: 0.13 # Slightly reduced from 0.14 to fine-tune its generalist multitask support. |
|
density: 0.65 # Increased from 0.63 to preserve more of its diverse capabilities. |
|
- model: sometimesanotion/Qwen2.5-14B-Vimarckoso |
|
parameters: |
|
weight: 0.11 # Slightly reduced from 0.12 to balance its multi-step reasoning contributions. |
|
density: 0.62 # Increased from 0.6 to retain more of its specialized reasoning strengths. |
|
- model: allknowingroger/QwenSlerp6-14B |
|
parameters: |
|
weight: 0.09 # Slightly reduced from 0.1 to refine its contextual reasoning contributions. |
|
density: 0.65 # Increased from 0.62 to preserve more of its capabilities in MMLU-PRO and contextual tasks. |
|
``` |
|
|
