đ ZYH-LLM-Qwen2.5-14B-V4
*ZYH-LLM-Qwen2.5-14B-V4 increases the proportion of the R1 distillation model in the model merging recipe while maintaining the model's instruction-following ability and *general capabilities.

đ Quick Start
The fifth-generation model of ZYH-LLM-Qwen2.5 has been released! Check it out here.
⨠Features
- Increase the proportion of the R1 distillation model in the model merging recipe.
- Maintain the model's instruction-following ability and general capabilities.
- Improve the calculation accuracy and inference ability of the model without reducing the general capabilities of the instruction model.
đ Documentation
Merge Template
merge_method: model_stock
base_model: Instruction Model
models:
- model: Instruction Fine-tuning Model 1
- model: Instruction Fine-tuning Model 2
- model: Inference Fine-tuning Model 1
- model: Inference Fine-tuning Model 2
dtype: bfloat16
tokenizer_source: base
int8_mask: true
normalize: true
Using the above template for merging can improve the calculation accuracy and inference ability of the model without reducing the general capabilities of the instruction model. ZYH-LLM-Qwen2.5-V4 used this template during the model merging process.
Detailed results can be found here
Property |
Details |
Avg. |
43.14 |
IFEval (0-Shot) |
83.65 |
BBH (3-Shot) |
50.27 |
MATH Lvl 5 (4-Shot) |
53.93 |
GPQA (0-shot) |
8.61 |
MuSR (0-shot) |
15.66 |
MMLU-PRO (5-shot) |
46.71 |
Model Merging Stages
First stage
Create four different instruction models and code model.
models:
- model: Qwen/Qwen2.5-14B-Instruct
parameters:
density: 1
weight: 1
lambda: 0.9
- model: Qwen/Qwen2.5-14B-Instruct-1M
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
tokenizer_source: base
name: Qwen2.5-14B-della-base
models:
- model: Qwen/Qwen2.5-14B-Instruct
parameters:
density: 1
weight: 1
lambda: 0.9
- model: Qwen/Qwen2.5-14B-Instruct-1M
parameters:
density: 1
weight: 1
lambda: 0.9
merge_method: della
base_model: arcee-ai/Virtuoso-Small-v2
parameters:
density: 1
weight: 1
lambda: 0.9
normalize: true
int8_mask: true
dtype: bfloat16
tokenizer_source: base
name: Qwen2.5-14B-della-v2
models:
- model: Qwen/Qwen2.5-14B-Instruct
parameters:
density: 1
weight: 1
lambda: 0.9
- model: Qwen/Qwen2.5-14B-Instruct-1M
parameters:
density: 1
weight: 1
lambda: 0.9
merge_method: della
base_model: arcee-ai/SuperNova-Medius
parameters:
density: 1
weight: 1
lambda: 0.9
normalize: true
int8_mask: true
dtype: bfloat16
tokenizer_source: base
name: Qwen2.5-14B-della-Nova
models:
- model: Qwen/Qwen2.5-14B-Instruct
parameters:
density: 1
weight: 1
lambda: 0.9
- model: Qwen/Qwen2.5-14B-Instruct-1M
parameters:
density: 1
weight: 1
lambda: 0.9
merge_method: della
base_model: Azure99/Blossom-V6-14B
parameters:
density: 1
weight: 1
lambda: 0.9
normalize: true
int8_mask: true
dtype: bfloat16
tokenizer_source: base
name: Qwen2.5-14B-della-V6
models:
- model: Qwen/Qwen2.5-Coder-14B-Instruct
parameters:
density: 1
weight: 1
lambda: 0.9
merge_method: della
base_model: Qwen/Qwen2.5-Coder-14B
parameters:
density: 1
weight: 1
lambda: 0.9
normalize: true
int8_mask: true
dtype: bfloat16
tokenizer_source: base
name: Qwen2.5-Coder-14B-della
Second stage
Step 1
Create three instruction models with a bias towards reasoning by using templates.
merge_method: model_stock
base_model: Qwen2.5-14B-della-base
models:
- model: Qwen2.5-Coder-14B-della
- model: Qwen2.5-14B-della-v2
- model: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
- model: huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2
dtype: bfloat16
tokenizer_source: base
int8_mask: true
normalize: true
name: Qwen2.5-14B-mst-Coder
merge_method: model_stock
base_model: Qwen2.5-14B-della-base
models:
- model: Qwen2.5-14B-della-V6
- model: Qwen2.5-14B-della-v2
- model: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
- model: huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2
dtype: bfloat16
tokenizer_source: base
int8_mask: true
normalize: true
name: Qwen2.5-14B-mst-V6
merge_method: model_stock
base_model: Qwen2.5-14B-della-base
models:
- model: Qwen2.5-14B-della-Nova
- model: Qwen2.5-14B-della-v2
- model: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
- model: huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2
dtype: bfloat16
tokenizer_source: base
int8_mask: true
normalize: true
name: Qwen2.5-14B-mst-Nova
Step 2
Create a pure instruction model to restore the generality of the final model.
merge_method: model_stock
base_model: Qwen2.5-14B-della-base
models:
- model: Qwen2.5-14B-della-Nova
- model: Qwen2.5-14B-della-v2
- model: Qwen2.5-14B-della-V6
dtype: bfloat16
tokenizer_source: base
int8_mask: true
normalize: true
name: Qwen2.5-14B-mst-it
Third stage
Create a base model with a context of 1 million tokens.
merge_method: sce
models:
- model: Qwen/Qwen2.5-14B-Instruct-1M
- model: Qwen/Qwen2.5-14B
base_model: Qwen/Qwen2.5-14B-Instruct-1M
parameters:
select_topk: 1
dtype: bfloat16
tokenizer_source: base
normalize: true
int8_mask: true
name: Qwen2.5-14B-1M
models:
- model: Qwen/Qwen2.5-14B-Instruct
parameters:
density: 1
weight: 1
lambda: 0.9
- model: Qwen/Qwen2.5-14B-Instruct-1M
parameters:
density: 1
weight: 1
lambda: 0.9
merge_method: della
base_model: Qwen2.5-14B-1M
parameters:
density: 1
weight: 1
lambda: 0.9
normalize: true
int8_mask: true
dtype: bfloat16
tokenizer_source: base
name: Qwen2.5-14B-della-1M
Final stage
merge_method: model_stock
base_model: Qwen2.5-14B-della-1M
models:
- model: Qwen2.5-14B-mst-Coder
- model: Qwen2.5-14B-mst-V6
- model: Qwen2.5-14B-mst-Nova
- model: Qwen2.5-14B-mst-it
dtype: bfloat16
tokenizer_source: base
int8_mask: true
normalize: true
name: ZYH-LLM-Qwen2.5-14B-V4
đ License
This project is licensed under the Apache-2.0 license.