64 lines
1.4 KiB
Markdown
64 lines
1.4 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
- zh
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen2.5-14B
|
||
|
|
- Qwen/Qwen2.5-14B-Instruct
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- merge
|
||
|
|
---
|
||
|
|
|
||
|
|
|
||
|
|
Combine the three methods of **della**, **ties**, and **model stock** to merge the instruction model with the base model.
|
||
|
|
|
||
|
|
The aim is to solve the problems of **the decline in instruction-following ability** and **mathematical ability** caused by using only the ties merging method or only the della merging method.
|
||
|
|
|
||
|
|
```yaml
|
||
|
|
models:
|
||
|
|
- model: Qwen/Qwen2.5-14B-instruct
|
||
|
|
parameters:
|
||
|
|
density: 1
|
||
|
|
weight: 1
|
||
|
|
lambda: 0.9
|
||
|
|
merge_method: della
|
||
|
|
base_model: Qwen/Qwen2.5-14B
|
||
|
|
parameters:
|
||
|
|
density: 1
|
||
|
|
weight: 1
|
||
|
|
lambda: 0.9
|
||
|
|
normalize: true
|
||
|
|
int8_mask: true
|
||
|
|
dtype: bfloat16
|
||
|
|
name: Qwen2.5-14B-della
|
||
|
|
```
|
||
|
|
```yaml
|
||
|
|
models:
|
||
|
|
- model: Qwen/Qwen2.5-14B-instruct
|
||
|
|
parameters:
|
||
|
|
density: 1
|
||
|
|
weight: 1
|
||
|
|
merge_method: ties
|
||
|
|
base_model: Qwen/Qwen2.5-14B
|
||
|
|
parameters:
|
||
|
|
density: 1
|
||
|
|
weight: 1
|
||
|
|
normalize: true
|
||
|
|
int8_mask: true
|
||
|
|
dtype: bfloat16
|
||
|
|
name: Qwen2.5-14B-ties
|
||
|
|
```
|
||
|
|
```yaml
|
||
|
|
merge_method: model_stock
|
||
|
|
base_model: Qwen/Qwen2.5-14B-instruct
|
||
|
|
models:
|
||
|
|
- model: Qwen2.5-14B-della
|
||
|
|
- model: Qwen2.5-14B-ties
|
||
|
|
dtype: bfloat16
|
||
|
|
tokenizer_source: base
|
||
|
|
int8_mask: true
|
||
|
|
normalize: true
|
||
|
|
name: Qwen2.5-14B-it-restore
|
||
|
|
```
|