A
mergekit model_stock merge made with the aim of improving WizardLM-2-8x22B.
The resulting model suppresses WizardLM-2-8x22B's overly flowery and positive writing style whilst retaining useful features such as CoT. Extremely coherant even at long contexts, and benched above the WLM-2 base in intelligence tests.
Use vicuna prompt as per WizardLM-2-8x22B base.
1models:
2 - model: alpindale/WizardLM-2-8x22B
3 - model: openbmb/Eurux-8x22b-kto
4 - model: openbmb/Eurux-8x22b-nca
5 - model: mistralai/Mixtral-8x22B-v0.1
6 - model: migtissera/Tess-2.0-Mixtral-8x22B
7 - model: fireworks-ai/mixtral-8x22b-instruct-oh
8base_model: alpindale/WizardLM-2-8x22B
9merge_method: model_stock
10dtype: bfloat16
Likely won't be uploading the full weights myself due to bandwidth limitations.
