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 LatitudeGames/Harbinger-24B as a base.
The following models were included in the merge:
The following YAML configuration was used to produce this model:
base_model: LatitudeGames/Harbinger-24B
merge_method: dare_ties
dtype: bfloat16
models:
- model: LatitudeGames/Harbinger-24B
parameters:
weight: 0.45
density: 0.92
- model: sarvamai/sarvam-m
parameters:
weight: 0.30
density: 0.86
- model: aixonlab/Eurydice-24b-v3.5
parameters:
weight: 0.25
density: 0.86
chat_template: auto
parameters:
normalize: true
int8_mask: true
tolerance: 1e-7