This is a version of Thespis which includes more than the usual amount of Assistant responses in the training data. Its good for acting as an assistant with personality. It may be ok for RP, feel free to experiment. It was also trained on DaringFortitude as opposed to the base Llama2 13b. This model works best with internet style RP using standard markup with asterisks surrounding actions and no quotes around dialogue. Currently in the middle of adding new features and working on making sure initial responses are strong. It uses a 40/60 Split of Assistant and RP data, and includes items from the below datasets: Pure-Dove Dataset Claude Multiround 30k No Robots OpenOrcaSlim * Augmental Dataset As a result of adding this additional Assistant data, the model shows greatly improved instruction following and coherence during multiturn chats. You can either bake usernames into the prompt directly for ease of use or programatically add them if running through the API to use as a chatbot.
Модальности:
Генерация текста
Задача: Генерация текста
Автор: cgato
Теги: llama, not-for-all-audiences, text-generation-inference, endpoints_compatible
Лайков: 3 | Загрузок: 21
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
