36 lines
1.1 KiB
YAML
36 lines
1.1 KiB
YAML
# Reproduces lighteternal/Llama3-merge-biomed-8b
|
|
# Original: https://huggingface.co/lighteternal/Llama3-merge-biomed-8b
|
|
#
|
|
# DARE-TIES merge of Llama-3-8B-Instruct + Hermes-2-Pro + OpenBioLLM
|
|
# The first two model entries share the base_model identity (delta=0).
|
|
# Only Hermes-2-Pro and OpenBioLLM contribute non-zero task vectors.
|
|
|
|
merge_method: dare_ties
|
|
# NousResearch mirror = identical weights to meta-llama/Meta-Llama-3-8B-Instruct
|
|
base_model: /tmp/hf_merge_ingredients/NousResearch/Meta-Llama-3-8B-Instruct
|
|
|
|
models:
|
|
- model: /tmp/hf_merge_ingredients/NousResearch/Meta-Llama-3-8B-Instruct
|
|
parameters:
|
|
density: 1.0
|
|
weight: 1.0
|
|
|
|
- model: /tmp/hf_merge_ingredients/NousResearch/Meta-Llama-3-8B-Instruct
|
|
parameters:
|
|
density: 0.60
|
|
weight: 0.5
|
|
|
|
- model: /tmp/hf_merge_ingredients/NousResearch/Hermes-2-Pro-Llama-3-8B
|
|
parameters:
|
|
density: 0.55
|
|
weight: 0.1
|
|
|
|
- model: /tmp/hf_merge_ingredients/aaditya/Llama3-OpenBioLLM-8B
|
|
parameters:
|
|
density: 0.55
|
|
weight: 0.4
|
|
|
|
parameters:
|
|
int8_mask: true
|
|
dtype: bfloat16
|