metadata
language:
- en
MergeMonster
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 occuring 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.