This is a merge of pre-trained language models created using
mergekit.
This model was merged using the SLERP merge method.
1models:
2 - model: djuna/Q2.5-Veltha-14B
3 - model: sometimesanotion/Qwen2.5-14B-Vimarckoso-v3
4merge_method: slerp
5base_model: djuna/Q2.5-Veltha-14B
6dtype: bfloat16
7parameters:
8 t: [0, 0.2, 0.8, 0.1, 0] # Skewed towards the second model
9
10regularization:
11 - method: gradient_penalty
12 scale: 0.07 # Increased for stronger regularization
13 - method: weight_clipping
14 clip_range: [-0.2, 0.2] # Widened range
15 - method: random_noise
16 scale: 0.005 # Reduced to avoid excessive noise
17 - method: attention_dropout
18 scale: 0.03 # Increased dropout
19
20postprocessing:
21 - operation: entropy_regularization
22 scale: 0.07 # Increased for stronger effect
23 - operation: non_linear_scaling
24 parameters:
25 function: gelu # Changed to GELU for potentially better performance
26 - operation: sharpening
27 intensity: 0.7 # Increased sharpening
28 - operation: gaussian_smoothing
29 sigma: 0.2 # Reduced smoothing for more detail
30 - operation: normalize
31 - operation: dynamic_scaling
32 scale_range: [0.97, 1.03] # Slightly wider range
33 - operation: smoothing
34 parameters:
35 adaptive: true
36 range: [0.97, 1.03] # Slightly wider range
37 kernel_size: 5 # Increased kernel size for more aggressive smoothing
