|
--- |
|
base_model: |
|
- CultriX/Qwen2.5-14B-Brocav7 |
|
- CultriX/Qwen2.5-14B-Emerged |
|
- sometimesanotion/Lamarck-14B-v0.6 |
|
- djuna/Q2.5-Veltha-14B-0.5 |
|
- allknowingroger/QwenSlerp6-14B |
|
- CultriX/SeQwence-14B-EvolMerge |
|
- hotmailuser/QwenSlerp2-14B |
|
- CultriX/Qwen2.5-14B-Hyperionv3 |
|
- CultriX/Qwen2.5-14B-Wernickev3 |
|
- qingy2024/Fusion4-14B-Instruct |
|
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/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: |
|
* [CultriX/Qwen2.5-14B-Brocav7](https://huggingface.co/CultriX/Qwen2.5-14B-Brocav7) |
|
* [CultriX/Qwen2.5-14B-Emerged](https://huggingface.co/CultriX/Qwen2.5-14B-Emerged) |
|
* [sometimesanotion/Lamarck-14B-v0.6](https://huggingface.co/sometimesanotion/Lamarck-14B-v0.6) |
|
* [djuna/Q2.5-Veltha-14B-0.5](https://huggingface.co/djuna/Q2.5-Veltha-14B-0.5) |
|
* [allknowingroger/QwenSlerp6-14B](https://huggingface.co/allknowingroger/QwenSlerp6-14B) |
|
* [CultriX/SeQwence-14B-EvolMerge](https://huggingface.co/CultriX/SeQwence-14B-EvolMerge) |
|
* [hotmailuser/QwenSlerp2-14B](https://huggingface.co/hotmailuser/QwenSlerp2-14B) |
|
* [CultriX/Qwen2.5-14B-Hyperionv3](https://huggingface.co/CultriX/Qwen2.5-14B-Hyperionv3) |
|
* [qingy2024/Fusion4-14B-Instruct](https://huggingface.co/qingy2024/Fusion4-14B-Instruct) |
|
|
|
### Configuration |
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
```yaml |
|
merge_method: dare_ties # Merge method for dynamic, task-aware parameter blending. |
|
base_model: CultriX/Qwen2.5-14B-Wernickev3 # Main backbone for parameter alignment. |
|
dtype: bfloat16 # Efficient precision for memory usage. |
|
out_dtype: bfloat16 # Output data type to maintain consistency and efficiency. |
|
|
|
parameters: |
|
epsilon: 0.010 # Fine-tuned scaling for precise parameter adjustments. |
|
lambda: 2.0 # Emphasizes high-impact parameters for improved task performance. |
|
normalize: true # Ensures parameter normalization for stability during merging. |
|
rescale: true # Rescales parameters across models for better integration. |
|
int8_mask: false # Disables int8 masking to preserve full precision. |
|
|
|
adaptive_merge_parameters: |
|
task_weights: # Weight prioritization for tasks. |
|
tinyArc: 1.6 # Balanced focus on logical reasoning. |
|
tinyHellaswag: 1.5 # Moderate priority for contextual reasoning. |
|
tinyMMLU: 1.8 # High priority for multi-domain knowledge tasks. |
|
tinyTruthfulQA: 2.2 # High emphasis on factual QA accuracy. |
|
tinyTruthfulQA_mc1: 1.8 # High priority for multiple-choice factual QA. |
|
tinyWinogrande: 1.75 # Moderate priority for contextual reasoning tasks. |
|
IFEval: 2.5 # Maximum priority for instruction-following tasks. |
|
BBH: 2.2 # High priority for complex reasoning tasks. |
|
MATH: 2.8 # Maximum priority for mathematical reasoning. |
|
GPQA: 2.2 # Balanced focus on graduate-level QA tasks. |
|
MUSR: 2.2 # High priority for multi-step reasoning. |
|
MMLU-PRO: 2.0 # High priority for multitask, domain-specific knowledge. |
|
smoothing_factor: 0.03 # Precise blending of task-specific contributions. |
|
|
|
gradient_clipping: # Gradient clipping for stability during merging. |
|
CultriX/Qwen2.5-14B-Wernickev3: 0.89 # Stability for the base model. |
|
djuna/Q2.5-Veltha-14B-0.5: 0.91 # Stability for reasoning contributions. |
|
CultriX/SeQwence-14B-EvolMerge: 0.87 # Stabilized for multitask performance. |
|
qingy2024/Fusion4-14B-Instruct: 0.93 # High stability for mathematical reasoning. |
|
CultriX/Qwen2.5-14B-Emerged: 0.89 # Stability for multitask contributions. |
|
sometimesanotion/Lamarck-14B-v0.6: 0.89 # Stability for multi-step reasoning. |
|
allknowingroger/QwenSlerp6-14B: 0.90 # Stability for general reasoning and multitask tasks. |
|
hotmailuser/QwenSlerp2-14B: 0.91 # Stabilized for instruction following. |
|
CultriX/Qwen2.5-14B-Hyperionv3: 0.90 # Stability for this model's general performance. |
|
CultriX/Qwen2.5-14B-Brocav7: 0.90 # Stability for specific task contributions. |
|
|
|
models: # Definition of models and their weights/densities. |
|
- model: CultriX/Qwen2.5-14B-Wernickev3 # Base generalist model. |
|
parameters: |
|
weight: 0.28 # Balanced weight for a strong backbone. |
|
density: 0.78 # Slightly reduced to balance smaller contributors. |
|
|
|
- model: djuna/Q2.5-Veltha-14B-0.5 # Reasoning-focused model. |
|
parameters: |
|
weight: 0.27 # Slightly reduced for better balance. |
|
density: 0.77 # Balanced density to ensure nuanced reasoning contributions. |
|
|
|
- model: allknowingroger/QwenSlerp6-14B # Strong multitask performer. |
|
parameters: |
|
weight: 0.15 # Balanced weight for generalist capabilities. |
|
density: 0.76 # Balanced density to maintain stable contributions. |
|
|
|
- model: hotmailuser/QwenSlerp2-14B # High IFEval performer. |
|
parameters: |
|
weight: 0.12 # Maintains stable contributions for instruction-following tasks. |
|
density: 0.70 # Increased density to enhance integration. |
|
|
|
- model: CultriX/Qwen2.5-14B-Hyperionv3 # Generalist model with solid performance. |
|
parameters: |
|
weight: 0.10 # Increased for balanced general contributions. |
|
density: 0.75 # Balanced density for stable integration. |
|
|
|
- model: CultriX/Qwen2.5-14B-Brocav7 # Model for specific tasks like reasoning. |
|
parameters: |
|
weight: 0.10 # Increased weight to strengthen specific contributions. |
|
density: 0.76 # Increased density for better parameter preservation. |
|
|
|
- model: CultriX/SeQwence-14B-EvolMerge # Multitask generalist. |
|
parameters: |
|
weight: 0.08 # Balanced weight for broader coverage. |
|
density: 0.68 # Slight increase for better integration. |
|
|
|
- model: qingy2024/Fusion4-14B-Instruct # Specialist in mathematical reasoning. |
|
parameters: |
|
weight: 0.08 # Balanced weight for MATH tasks. |
|
density: 0.78 # Increased density to enhance task-specific integration. |
|
|
|
- model: CultriX/Qwen2.5-14B-Emerged # General multitask model. |
|
parameters: |
|
weight: 0.08 # Balanced for multitask contributions. |
|
density: 0.72 # Increased density for better parameter alignment. |
|
|
|
- model: sometimesanotion/Lamarck-14B-v0.6 # Multi-step reasoning focus. |
|
parameters: |
|
weight: 0.05 # Slightly increased to improve its contributions. |
|
density: 0.65 # Increased for better parameter blending. |
|
``` |
|
|
