| # MergeKit Configuration for merging alignment-DPO with roleplay capabilities | |
| merge_method: task_arithmetic | |
| # Define the base model (your alignment model) | |
| base_model: | |
| model: AlexCuadron/chai-ddpo # Path to your DPO-trained alignment model | |
| # If your model is in Hugging Face format, you could use: | |
| model_type: huggingface | |
| # If it's a local model, just use the path: | |
| # model: /path/to/alex_dpo | |
| # Models to merge with the base model | |
| models: | |
| - model: Nitral-AI/Community_Request-02-12B | |
| # Modify if needed: | |
| model_type: huggingface | |
| parameters: | |
| weight: 0.3 # Adjust between 0.2-0.4 based on how much roleplay vs alignment you want | |
| # Output settings | |
| dtype: bfloat16 # Or bfloat16 if your GPU supports it | |
| output_dir: ./aligned_roleplay_model | |
| # Optional: include parameters you want to skip or include specifically | |
| # parameters_settings: | |
| # include: [".*"] # Include all parameters, or specify patterns to include | |
| # exclude: [] # No exclusions, or specify patterns to exclude | |
| # Add this if you want to save disk space | |
| output_optimizations: | |
| safetensors: true # Save in safetensors format (more secure and efficient) | |
| # set to false if you want to keep intermediate models | |
| skip_saving_intermediate_models: true |