CloudGoat / Mephisto-4B-v2

Mephisto-4B-v2

This is an agentic model created by merging multiple fine-tuned models based on Qwen3.5-4B using a version of Mergekit forked to support Qwen3.5. Compared to Mephisto-4B, v2 involved a more meaningful merging process; rather than completing it in a single step, a multi-stage merging approach was used.

Merge Details

Merge Method

1. Merge Jackrong/Qwopus3.5-4B-v3 and Jackrong/Qwopus3.5-4B-Coder using the SLERP method → ​​Create Model 1
2. Merge BAAI/AREX-Turbo and InternScience/Agents-A1-4B using the SLERP method → ​​Create Model 2
3. Merge Model 1 and Model 2 using DARE-TIES → This model

Configuration

I performed the merge according to these YAML files:]

Model 1 (Qwopus3.5-4B-Fusion)

slices:
  - sources:
      - model: Jackrong/Qwopus3.5-4B-v3
        layer_range: [0, 32]
      - model: Jackrong/Qwopus3.5-4B-Coder
        layer_range: [0, 32]
merge_method: slerp
base_model: Jackrong/Qwopus3.5-4B-v3
parameters:
  t:
    - filter: self_attn
      value: [0.2, 0.3, 0.4, 0.5, 0.6, 0.6, 0.5]
    - filter: mlp
      value: [0.5, 0.5, 0.5, 0.5, 0.5]
    - value: 0.5
dtype: bfloat16

Model 2 (AREX-Turbo + Agents-A1-4B)

slices:
  - sources:
      - model: BAAI/AREX-Turbo
        layer_range: [0, 32]
        weight: 0.5
      - model: InternScience/Agents-A1-4B
        layer_range: [0, 32]
        weight: 0.5
merge_method: slerp
base_model: BAAI/AREX-Turbo
parameters:
  t:
    - value: 0.5
dtype: bfloat16

Mephisto-4B-v2

merge_method: dare_ties
base_model: /path/to/Qwen/Qwen3.5-4B

models:
  - model: /home/CloudGoat/llms_merge/Qwents-4B
    parameters:
      weight: 1.0
      density: 1.0

  - model: /home/CloudGoat/llms_merge/Qwopus3.5-4B-Fusion
    parameters:
      weight: 0.7
      density: 0.3

parameters:
  int8_mask: true
  normalize: false

dtype: bfloat16

Benchmarks

In progress