This model started as a block-diagonal frankenllama merge between MythoMax13B and Enterredaas33B However due to some anomaly likely caused by the novel methods used by MythoMax I was unable to initiate the LoRA training needed to bring the resulting model back to order. Being a Chronorctypus-Limarobormes enjoyer I decided to look further into the TIES-merging that it utilizes- as cited in the arXiv paper: Resolving Interference When Merging Models I used llama2-22b as the base model upon which I merged the MythoMax/Enterredaas frankenmerge, Dendrite-II and Bacchus Suffice it to say due to being made of merges of merges Yousei-22B has the DNA of many different models in it, although the Llama-2-chat DNA from Dendrite-II and the DNA from Mytho I no longer use simple-proxy-for-tavern as part of my testing stack and instead test on a combination of the kobold-lite UI and the latest version of silly-tavern (now with more advanced prompt formatting options) When using the llama-2 prompt format it tends to produce a lot of refusals when requesting mature or controversial content. When using an alpaca-style format you don’t get as many refusals, however…
Модальности:
Генерация текста
Задача: Генерация текста
Автор: Envoid
Теги: llama, text-generation-inference, endpoints_compatible
Лайков: 2 | Загрузок: 54
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
