98 lines
2.2 KiB
Markdown
98 lines
2.2 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
- zh
|
||
|
|
base_model:
|
||
|
|
- Kwai-Klear/Klear-Reasoner-8B
|
||
|
|
- deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
|
||
|
|
- Qwen/Qwen3-8B
|
||
|
|
- AXCXEPT/Qwen3-EZO-8B-beta
|
||
|
|
- Qwen/Qwen3-8B-Base
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- merge
|
||
|
|
---
|
||
|
|
> *Enhance the performance of Qwen3-8B by merging powerful reasoning models without compromising the effectiveness of the \no_think tag!*
|
||
|
|
# *Model Highlights:*
|
||
|
|
|
||
|
|
- ***merge method**: `arcee_fusion` `nuslerp` `della`*
|
||
|
|
|
||
|
|
- ***precision**: `dtype: float32` `out_dtype: bfloat16`*
|
||
|
|
|
||
|
|
- ***Context length**: `40960`*
|
||
|
|
|
||
|
|
# *Parameter Settings:*
|
||
|
|
## *Thinking Mode:*
|
||
|
|
> [!NOTE]
|
||
|
|
> *`Temperature=0.6`, `TopP=0.95`, `TopK=20`,`MinP=0`.*
|
||
|
|
## *Non-Thinking Mode:*
|
||
|
|
> [!TIP]
|
||
|
|
> *`Temperature=0.7`, `TopP=0.8`, `TopK=20`,`MinP=0`.*
|
||
|
|
|
||
|
|
## *Step1: Merge Two Hybrid Models*
|
||
|
|
- *Leverage the advantages of the two hybrid models.*
|
||
|
|
```yaml
|
||
|
|
models:
|
||
|
|
- model: AXCXEPT/Qwen3-EZO-8B-beta
|
||
|
|
merge_method: arcee_fusion
|
||
|
|
base_model: Qwen/Qwen3-8B
|
||
|
|
dtype: float32
|
||
|
|
out_dtype: bfloat16
|
||
|
|
tokenizer_source: base
|
||
|
|
name: Qwen3-8B-fusion
|
||
|
|
```
|
||
|
|
## *Step2: Merge High-Performance Reasoning Models with Hybrid Models*
|
||
|
|
- *Maximize the proportion of reasoning models on the premise that the \no_think tag remains effective.*
|
||
|
|
```yaml
|
||
|
|
models:
|
||
|
|
- model: Qwen3-8B-fusion
|
||
|
|
parameters:
|
||
|
|
weight: 0.55
|
||
|
|
- model: deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
|
||
|
|
parameters:
|
||
|
|
weight: 0.45
|
||
|
|
merge_method: nuslerp
|
||
|
|
tokenizer_source: Qwen/Qwen3-8B
|
||
|
|
dtype: float32
|
||
|
|
out_dtype: bfloat16
|
||
|
|
name: Qwen3-8B-nuslerp1
|
||
|
|
```
|
||
|
|
```yaml
|
||
|
|
models:
|
||
|
|
- model: Qwen3-8B-fusion
|
||
|
|
parameters:
|
||
|
|
weight: 0.75
|
||
|
|
- model: Kwai-Klear/Klear-Reasoner-8B
|
||
|
|
parameters:
|
||
|
|
weight: 0.25
|
||
|
|
merge_method: nuslerp
|
||
|
|
tokenizer_source: Qwen/Qwen3-8B
|
||
|
|
dtype: float32
|
||
|
|
out_dtype: bfloat16
|
||
|
|
name: Qwen3-8B-nuslerp2
|
||
|
|
```
|
||
|
|
## *Step3: Unify the Enhanced Hybrid Modes*
|
||
|
|
- *Merge the two models into the base model using the della merging method to make the model more versatile and stable.*
|
||
|
|
- *We use the chat template of Qwen3-8B.*
|
||
|
|
```yaml
|
||
|
|
models:
|
||
|
|
- model: Qwen3-8B-nuslerp1
|
||
|
|
parameters:
|
||
|
|
density: 1
|
||
|
|
weight: 1
|
||
|
|
lambda: 0.9
|
||
|
|
- model: Qwen3-8B-nuslerp2
|
||
|
|
parameters:
|
||
|
|
density: 1
|
||
|
|
weight: 1
|
||
|
|
lambda: 0.9
|
||
|
|
merge_method: della
|
||
|
|
base_model: Qwen/Qwen3-8B-Base
|
||
|
|
dtype: float32
|
||
|
|
out_dtype: bfloat16
|
||
|
|
name: Qwen3-8B-YOYO-V2-Hybrid
|
||
|
|
```
|
||
|
|
|
||
|
|
|