DELLA-Merging: Reducing Interference in Model Merging through Magnitude-Based Sampling
Paper
•
2406.11617
•
Published
•
8
This is a merge of pre-trained language models created using mergekit.
This model was merged using the DELLA merge method using /workspace/prototype-0.4x228 as a base.
The following models were included in the merge:
The following YAML configuration was used to produce this model:
models:
- model: /workspace/prototype-0.4x234
parameters:
weight: 0.2
density: 0.4
epsilon: 0.15
- model: /workspace/prototype-0.4x233
parameters:
weight: 0.2
density: 0.4
epsilon: 0.15
- model: /workspace/prototype-0.4x232
parameters:
weight: 0.2
density: 0.4
epsilon: 0.15
- model: /workspace/prototype-0.4x231
parameters:
weight: 0.2
density: 0.4
epsilon: 0.15
- model: /workspace/prototype-0.4x228
parameters:
weight: 0.2
density: 0.4
epsilon: 0.1
base_model: /workspace/prototype-0.4x228
merge_method: della
tokenizer:
source: base
parameters:
normalize: false
lambda: 1.1
pad_to_multiple_of: 8
int8_mask: true
chat_template: llama3
dtype: bfloat16