How you fine tunned this model?
#1
by
celsowm
- opened
Hi!
I read on reddit that this model used a new technique to inject new domain knowledge.
Could you explain it?
You can easily create additional layers using mergekit(https://github.com/arcee-ai/mergekit). Use the following settings It is a simple task to unfreeze and train only the added layer.
slices:
- sources:
- model: meta-llama/Meta-Llama-3-8B-Instruct
layer_range: [0, 20]
- sources:
- model: meta-llama/Meta-Llama-3-8B-Instruct
layer_range: [12, 32]
merge_method: passthrough
dtype: bfloat16