I'm going to compare DARE merges using this (mostly vanilla, alpaca-tinted) 20b model vs using Harmonia. | |
slices: | |
- sources: | |
- model: athirdpath/alpaca-2-13b-english_full-model | |
- | |
layer_range: [0, 16] | |
- sources: | |
- model: TheBloke/Llama-2-13B-fp16 | |
- | |
layer_range: [8, 24] | |
- sources: | |
- model: athirdpath/alpaca-2-13b-english_full-model | |
- | |
layer_range: [17, 32] | |
- sources: | |
- model: TheBloke/Llama-2-13B-fp16 | |
- | |
layer_range: [25, 40] | |
merge_method: passthrough | |
dtype: float16 |