Resolving Interference When Merging Models
Paper
•
2306.01708
•
Published
•
15
This is a merge of pre-trained language models created using mergekit.
This model was merged using the TIES merge method using mistralai/Mistral-7B-Instruct-v0.2 as a base.
The following models were included in the merge:
The following YAML configuration was used to produce this model:
models:
- model: mistralai/Mistral-7B-Instruct-v0.2
# Базовая модель, не меняем
parameters:
density: 1.0
weight: 1.0
- model: WizardLM/WizardMath-7B-V1.1
# Эксперт подмешивается
parameters:
density: 0.5 # DARE: берем только 50% самых важных изменений
weight: 0.5 # Вес влияния
merge_method: ties
base_model: mistralai/Mistral-7B-Instruct-v0.2
parameters:
normalize: true
int8_mask: true
dtype: float16