This is a merge of pre-trained language models created using
mergekit.
This model was merged using the
Task Arithmetic merge method using
Nitral-AI/Captain-Eris_Violet-V0.420-12B as a base.
1# MergeKit Configuration for merging alignment-DPO with roleplay capabilities
2merge_method: task_arithmetic
3
4# Define the base model (your alignment model)
5base_model:
6 model: Nitral-AI/Captain-Eris_Violet-V0.420-12B # Nitral-AI/Community_Request-02-12B # Path to your DPO-trained alignment model
7 # If your model is in Hugging Face format, you could use:
8 model_type: huggingface
9 # If it's a local model, just use the path:
10 # model: /path/to/alex_dpo
11
12# Models to merge with the base model
13models:
14 - model: AlexCuadron/chai-ddpo
15 # Modify if needed:
16 model_type: huggingface
17 parameters:
18 weight: 0.2 # Adjust between 0.2-0.4 based on how much roleplay vs alignment you want
19
20# Output settings
21dtype: bfloat16 # Or bfloat16 if your GPU supports it
22output_dir: ./aligned_roleplay_model
23# Optional: include parameters you want to skip or include specifically
24# parameters_settings:
25# include: [".*"] # Include all parameters, or specify patterns to include
26# exclude: [] # No exclusions, or specify patterns to exclude
27
28# Add this if you want to save disk space
29output_optimizations:
30 safetensors: true # Save in safetensors format (more secure and efficient)
31 # set to false if you want to keep intermediate models
32 skip_saving_intermediate_models: true
