Llama 3 Experiments
Collection
My merging experiments for the Llama-3 series of models.
•
6 items
•
Updated
This is a naive (passthrough) self-merge of two NousResearch/Hermes-2-Pro-Llama-3-8B, using mergekit. The merge config was inpired by mlabonne/Meta-Llama-3-120B-Instruct.
This gives better (and actually usable) results compared to my previous merges. I can clearly see a rise in capability as it responded with a more thought out answer to my function calling prompt, but failed to adhere to the format and instructions fully. Will read more and look at other merge methods.
The GGUF version(s) can be found here:
This model was merged using the passthrough merge method.
The following models were included in the merge:
The following YAML configuration was used to produce this model:
slices:
- sources:
- layer_range: [0, 8]
model: NousResearch/Hermes-2-Pro-Llama-3-8B
- sources:
- layer_range: [4, 12]
model: NousResearch/Hermes-2-Pro-Llama-3-8B
- sources:
- layer_range: [8, 16]
model: NousResearch/Hermes-2-Pro-Llama-3-8B
- sources:
- layer_range: [12, 20]
model: NousResearch/Hermes-2-Pro-Llama-3-8B
- sources:
- layer_range: [16, 24]
model: NousResearch/Hermes-2-Pro-Llama-3-8B
- sources:
- layer_range: [20, 28]
model: NousResearch/Hermes-2-Pro-Llama-3-8B
- sources:
- layer_range: [24, 32]
model: NousResearch/Hermes-2-Pro-Llama-3-8B
merge_method: passthrough
dtype: float16