Ateron/Gemma-4-Dark-Thoughts-31B

🚬 Dark-Thoughts

"Shady corners of mind"

🎥 Overview

This is another merge made for role-play and creative writing. I used latest models that I found one of the best in my own hours of testings. It can pick up easily system prompts and follow direction of the story, without passiveness.

💿 Mergekit Configuration

Below is the exact mergekit_config.yml recipe used to synthesize this model:

Phase 1: Spark

Ignition of creativity with MeroMero V2 and Dark-Scarlett V2

dare_ties Recipe 1

merge_method: dare_ties
base_model: F:\AI\Merge\Gemma-4-it
tokenizer_source: base
dtype: bfloat16
parameters:
  lambda: 1.0
models:
  - model: F:\AI\Merge\Gemma-4-MeroMero-V2
    parameters:
      density: 0.50
      weight:
        - filter: "model.layers.4."
          value: 0.70
        - filter: "model.layers.9."
          value: 0.70
        - filter: "model.layers.14."
          value: 0.70
        - filter: "model.layers.19."
          value: 0.70
        - filter: "model.layers.24."
          value: 0.70
        - filter: "model.layers.29."
          value: 0.70
        - filter: "model.layers.34."
          value: 0.70
        - filter: "model.layers.39."
          value: 0.60
        - filter: "model.layers.44."
          value: 0.50
        - filter: "model.layers.49."
          value: 0.50
        - filter: "model.layers.54."
          value: 0.50
        - filter: "model.layers.59."
          value: 0.50
        - value: [0.30, 0.30, 0.30, 0.30, 0.30]

  - model: F:\AI\Merge\Gemma-4-Dark-Scarlett-V2
    parameters:
      density: 0.50
      weight:
        - filter: "model.layers.4."
          value: 0.30
        - filter: "model.layers.9."
          value: 0.30
        - filter: "model.layers.14."
          value: 0.30
        - filter: "model.layers.19."
          value: 0.30
        - filter: "model.layers.24."
          value: 0.30
        - filter: "model.layers.29."
          value: 0.30
        - filter: "model.layers.34."
          value: 0.30
        - filter: "model.layers.39."
          value: 0.40
        - filter: "model.layers.44."
          value: 0.50
        - filter: "model.layers.49."
          value: 0.50
        - filter: "model.layers.54."
          value: 0.50
        - filter: "model.layers.59."
          value: 0.50
        - value: [0.30, 0.30, 0.30, 0.30, 0.30]

Phase 2: Form

Polisher of thoughts from Scotoma v2

dare_ties Recipe 2

merge_method: dare_ties
base_model: F:\AI\Merge\Gemma-4-it
tokenizer_source: base
dtype: bfloat16
parameters:
  lambda: 1.0
models:
  - model: F:\AI\Merge\Gemma-4-Scotoma-V2
    parameters:
      density: 0.50
      weight:
        - filter: "model.layers.4."
          value: 0.80
        - filter: "model.layers.9."
          value: 0.80
        - filter: "model.layers.14."
          value: 0.80
        - filter: "model.layers.19."
          value: 0.80
        - filter: "model.layers.24."
          value: 0.80
        - filter: "model.layers.29."
          value: 0.80
        - filter: "model.layers.34."
          value: 0.70
        - filter: "model.layers.39."
          value: 0.60
        - filter: "model.layers.44."
          value: 0.50
        - filter: "model.layers.49."
          value: 0.50
        - filter: "model.layers.54."
          value: 0.50
        - filter: "model.layers.59."
          value: 0.50
        - value: [0.70, 0.70, 0.70, 0.70, 0.70]

  - model: F:\AI\Merge\DarkMero
    parameters:
      density: 0.50
      weight:
        - filter: "model.layers.4."
          value: 0.20
        - filter: "model.layers.9."
          value: 0.20
        - filter: "model.layers.14."
          value: 0.20
        - filter: "model.layers.19."
          value: 0.20
        - filter: "model.layers.24."
          value: 0.20
        - filter: "model.layers.29."
          value: 0.20
        - filter: "model.layers.34."
          value: 0.30
        - filter: "model.layers.39."
          value: 0.40
        - filter: "model.layers.44."
          value: 0.50
        - filter: "model.layers.49."
          value: 0.50
        - filter: "model.layers.54."
          value: 0.50
        - filter: "model.layers.59."
          value: 0.50
        - value: [0.30, 0.30, 0.30, 0.30, 0.30]

🤝 Special Thanks

• Google DeepMind: For providing the base model.
• The Open-Source Community: Creators and all their fine-tuned models that were used in the merge.
• Mergekit Fork: Zerofata - For making mergekit work.
• To Nimbz: This cat for being a smart fella and assisting with advices.