L3-ColdBrew-OldSpice / mergekit_config.yml
Theros's picture
Upload folder using huggingface_hub
8eebbfd verified
raw
history blame contribute delete
624 Bytes
const_tag: &scale_factor 0.7071067812 # 1/sqrt(2) scaling for stability
attenuate-env: &attenuated_env
parameters:
scale:
- filter: q_proj
value: *scale_factor
- filter: k_proj
value: *scale_factor
- value: 1.0
slices:
slices:
- sources:
- model: SvalTek/L3-ColdBrew-Astrid
layer_range: [0, 20]
- sources:
- model: SvalTek/L3-ColdBrew-Astrid
layer_range: [12, 20]
<<: *attenuated_env
- sources:
- model: SvalTek/L3-ColdBrew-SpicyReflect
layer_range: [20, 32] # Higher reasoning layers
merge_method: passthrough
dtype: bfloat16