Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch
Paper
•
2311.03099
•
Published
•
30
This is a merge of pre-trained language models created using mergekit.
This model was merged using the DARE TIES merge method using /workspace/cache/models--bruhzair--ignore-merge-17/snapshots/bd0af76a6bc4d9ae4bab5fa6b50e6545e6f3fd4f as a base.
The following models were included in the merge:
The following YAML configuration was used to produce this model:
base_model: /workspace/cache/models--bruhzair--ignore-merge-17/snapshots/bd0af76a6bc4d9ae4bab5fa6b50e6545e6f3fd4f
dtype: float32
merge_method: dare_ties
modules:
default:
slices:
- sources:
- layer_range: [0, 120]
model: /workspace/cache/models--bruhzair--ignore-merge-19/snapshots/57e2fc7118091e0706844a2e88d4c911d94e1e52
parameters:
density: 0.2
weight: 0.5
- layer_range: [0, 120]
model: /workspace/cache/models--bruhzair--ignore-merge-10-nemo/snapshots/bc2bc1ac38a2d05d5d32a31dd24d109bcc37c64c
parameters:
density: 0.2
weight: 0.5
- layer_range: [0, 120]
model: /workspace/cache/models--bruhzair--ignore-merge-17/snapshots/bd0af76a6bc4d9ae4bab5fa6b50e6545e6f3fd4f
out_dtype: bfloat16
parameters:
normalize: 1.0
tokenizer:
source: base