metadata
base_model:
- unsloth/Mistral-Small-Instruct-2409
library_name: transformers
tags:
- mergekit
- merge
MB-Zephyria-45b [EXPERIMENTAL]
Model Information
Base Model: unsloth/Mistral-Small-Instruct-2409
Strategy: Modified Balanced Approach with Extended Duplication
Total Layers: 55
Duplication Start: Layer 19 (34.5% of model)
Duplicated Layers: 30 (54.5% of model)
Unique Final Layers: 7 (11% of model)
Model Characteristics
- Models down_proj and o_proj layers have been nulled and will require healing
- Extends duplication further into later layers compared to the Balanced Approach
- Aims to enhance both understanding and creativity
- Maintains substantial unique initial layers for foundational processing
- Potentially suitable for complex reasoning and generative tasks
Configuration Visualization
[ Unique ][ Duplicated ][Unique]
0 ----------- 18 19 ------------------- 48 49 --- 54
34.5% 54.5% 11%