Qwen2.5-32B-YOYO-V2

The YOYO Second Generation 32B Model is Released!

Highlights:

1. Using the Karcher merging method.

2. Integrating high-performance 32B models from the open-source community.

First stage:

Make a code model:

models:
  - model: Qwen/Qwen2.5-Coder-32B-instruct
    parameters:
      density: 1 
      weight: 1
      lambda: 0.9
merge_method: della
base_model: Qwen/Qwen2.5-Coder-32B
parameters:
  density: 1
  weight: 1
  lambda: 0.9
  normalize: true
  int8_mask: true
dtype: bfloat16
name: YOYO-AI/Qwen2.5-Coder-32B-YOYO

Second stage:

Make an instruction model:

models:
  - model: YOYO-AI/Qwen2.5-Coder-32B-YOYO
  - model: Qwen/QwQ-32B
  - model: Skywork/Skywork-OR1-32B-Preview
  - model: deepcogito/cogito-v1-preview-qwen-32B
  - model: qihoo360/Light-R1-32B
  - model: AXCXEPT/EZO-Qwen2.5-32B-Instruct
  - model: fblgit/TheBeagle-v2beta-32B-MGS
  - model: tanliboy/lambda-qwen2.5-32b-dpo-test
  - model: Qwen/Qwen2.5-32B-Instruct
merge_method: karcher
base_model: Qwen/Qwen2.5-32B-Instruct
parameters:
  max_iter: 1000
normalize: true
int8_mask: true
tokenizer_source: base
dtype: float16
name: YOYO-AI/Qwen2.5-32B-YOYO-karcher

Third stage:

Make a base model:

models:
  - model: EVA-UNIT-01/EVA-Qwen2.5-32B-v0.2
  - model: Azure99/Blossom-V6-32B
  - model: arcee-ai/Virtuoso-Medium-v2
merge_method: karcher
base_model: Qwen/Qwen2.5-32B
parameters:
  max_iter: 1000
normalize: true
int8_mask: true
tokenizer_source: base
dtype: float16
name: YOYO-AI/Qwen2.5-32B-YOYO-karcher-base

Final stage:

models:
  - model: YOYO-AI/Qwen2.5-32B-YOYO-karcher
    parameters:
      density: 1
      weight: 1
      lambda: 0.9
merge_method: della
base_model: YOYO-AI/Qwen2.5-32B-YOYO-karcher-base
parameters:
  density: 1
  weight: 1
  lambda: 0.9
  normalize: true
  int8_mask: true
dtype: bfloat16
tokenizer_source: base
name: YOYO-AI/Qwen2.5-32B-YOYO-V2