gemma-4-31B-merge-experiment API & Inference Endpoint

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using MRockatansky/Gemma-4-31B-storymaxxed2 as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

yaml
# Model Merge Configuration for Creative Writing & Roleplay Gemma-4
# Combines: Storymaxxed (story writing), MeroMero (roleplay), Pantheon (reasoning)
# Merge Method: DARE TIES - Best for 3+ specialized models with distinct capabilities

merge_method: dare_ties

base_model: MRockatansky/Gemma-4-31B-storymaxxed2  # Using Storymaxxed as base

models:
  # Primary: Story Writing Specialist
  - model: MRockatansky/Gemma-4-31B-storymaxxed2
    parameters:
      weight: 0.4
      density: 0.7  # Keep 70% of task vector parameters

  # Secondary: Creative Roleplay Specialist  
  - model: zerofata/G4-MeroMero-31B
    parameters:
      weight: 0.3
      density: 0.7  # Keep 70% of task vector parameters

  # Tertiary: Reasoning Enhancement
  - model: Gryphe/Pantheon-Reasoning-31B-1.1
    parameters:
      weight: 0.3
      density: 0.7  # Keep 70% of task vector parameters

parameters:
  # Normalize weights across models
  normalize: true
  
  # Use int8 masks for memory efficiency (important for 31B models)
  int8_mask: true

# Output precision - bfloat16 recommended for Gemma-4
dtype: bfloat16

# Optional: Tokenizer configuration
# Using union to preserve any special tokens from all models
tokenizer:
  source: union

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using MRockatansky/Gemma-4-31B-storymaxxed2 as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

yaml
# Model Merge Configuration for Creative Writing & Roleplay Gemma-4
# Combines: Storymaxxed (story writing), MeroMero (roleplay), Pantheon (reasoning)
# Merge Method: DARE TIES - Best for 3+ specialized models with distinct capabilities

merge_method: dare_ties

base_model: MRockatansky/Gemma-4-31B-storymaxxed2  # Using Storymaxxed as base

models:
  # Primary: Story Writing Specialist
  - model: MRockatansky/Gemma-4-31B-storymaxxed2
    parameters:
      weight: 0.4
      density: 0.7  # Keep 70% of task vector parameters

  # Secondary: Creative Roleplay Specialist  
  - model: zerofata/G4-MeroMero-31B
    parameters:
      weight: 0.3
      density: 0.7  # Keep 70% of task vector parameters

  # Tertiary: Reasoning Enhancement
  - model: Gryphe/Pantheon-Reasoning-31B-1.1
    parameters:
      weight: 0.3
      density: 0.7  # Keep 70% of task vector parameters

parameters:
  # Normalize weights across models
  normalize: true
  
  # Use int8 masks for memory efficiency (important for 31B models)
  int8_mask: true

# Output precision - bfloat16 recommended for Gemma-4
dtype: bfloat16

# Optional: Tokenizer configuration
# Using union to preserve any special tokens from all models
tokenizer:
  source: union

gemma-4-31B-merge-experiment

README

Merge Details

Merge Method

Models Merged

Configuration

Explore FriendliAI today

README

Merge Details

Merge Method

Models Merged

Configuration