Views
No views yet
1merge_method: slerp
2base_model: "/mnt/c/Users/jdesa/finalround_8" # Path to the first model
3models:
4 - model: "/mnt/c/Users/jdesa/finalround_5" # Path to the second model
5
6# Interpolation parameters for SLERP
7parameters:
8 t:
9 - filter: self_attn # Interpolation values for self-attention layers
10 value: [0, 0.5, 0.7, 0.9, 1] # Gradients for SLERP interpolation
11 - filter: mlp # Interpolation values for MLP layers
12 value: [1, 0.7, 0.5, 0.3, 0] # Gradients for SLERP interpolation
13 - value: 0.5 # Default interpolation value for all other layers
14
15dtype: bfloat16 # Precision for merging
16tokenizer:
17 source: base # Use tokenizer from the base model
18