An unsupervised merging algorithm for Transformers-based language models, using a list of phrases (both good and bad) and a fully automated strategy that strives to decrease (or increase) the probability of these phrases occurring in the final merge.
This repo serves as a place to store various promising quantizations of merges produced by the Merge Monster.
Current variations
MergeMonster-7b - First model: Mistral 0.1, with the objective to reduce overused GPT style roleplaying phrases.
MergeMonster-13b - First model: MythoMax, with the objective to reduce overused GPT style roleplaying phrases.
MergeMonster-Decensored-7b - First model: Mistral 0.1, with the objective to reduce refusals on a global level.
MergeMonster-Decensored-13b - First model: MythoMax, with the objective to reduce refusals on a global level.
MergeMonster-WritingStyle-7b - First model: Mistral 0.1, with the singular goal to increase the chance of completing a nicely written roleplaying paragraph.
