Oh no, he's dumb too! I have a working hypothesis. Inverting and merging 20b Llama 2 models works quite well, evening out the gradients between slices. However, these 13b Mistrals seem to HATE it, I assume due to the unbalanced nature of my recipe. More study is required.
Recipe
merge_method: dare_ties
base_model: athirdpath/BigMistral-13b
model: athirdpath/CleverMage-Mistral-13b
weight: 0.60 / density: 0.35
model: athirdpath/CleverMage-Mistral-13b-INV
weight: 0.40 / density: 0.30
int8_mask: true
dtype: bfloat16