I’m just tinkering. All credit to the original creator: Undi. This model is a more traditionally quantized EXL2 compared to my usual «rpcal» versions. Llama-3-8B seems to get markedly dumber by using the «rpcal» method. In previous models, it was difficult to tell, but the margin of error increase from quantizing Llama-3-8B makes it obvious which method is better. I deleted the lower quants of rpcal because they are pretty dumb by comparison. I recommend this version, and this version only for this model. Lower quants are significantly worse. * This model: EXL2 @ 8.0 bpw. This is a merge of pre-trained language models created using mergekit. The new EVOLVE merge method was used (on MMLU specifically), see below for more information! Unholy was used for uncensoring, Roleplay Llama 3 for the DPO train he got on top, and LewdPlay for the… lewd side. This model was merged using the DARE TIES merge method using ./mergekit/inputmodels/Roleplay-Llama-3-8B213413727 as a base. The following models were included in the merge: ./mergekit/inputmodels/Llama-3-Unholy-8B-e41440388923 ./mergekit/inputmodels/Llama-3-LewdPlay-8B-e32981937066 The following YAML configuration was used to produce…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: zaq-hack
Теги: llama, not-for-all-audiences, nsfw, merge, conversational, text-generation-inference, endpoints_compatible, 8-bit
Лайков: 3 | Загрузок: 15
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
