Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch
Paper • 2311.03099 • Published • 32
This is a merge of pre-trained language models created using mergekit.
This model was merged using the DARE TIES merge method using Delta-Vector/Rei-12B as a base.
The following models were included in the merge:
The following YAML configuration was used to produce this model:
models:
- model: Delta-Vector/Rei-12B
parameters:
weight: 1
density: 0.85
- model: redrix/patricide-12B-Unslop-Mell
parameters:
weight: 0.72
density: 0.89
- model: IlyaGusev/saiga_nemo_12b
parameters:
weight: 0.18
density: 0.71
merge_method: dare_ties
base_model: Delta-Vector/Rei-12B
parameters:
density: 0.9
epsilon: 0.05
lambda: 1.1
dtype: bfloat16