Another experimental model, using mostly sythetic data generated by airoboros This models uses llama-2 chat format, rather than vicuna style user/assistant! This is a breaking change, although most inference systems support llama-2 chat templating. This is just one last release tweaking hyperparams, which seem to result in a higher quality model than 3.1 or 3.1.1 (and doens’t have the prompt issue of 3.1). A model built on airoboros-3.1 dataset, which is a continuation of airoboros-3.0 dataset with the following extras: More MathJSON, now ~17k items — math questions, prefixed with «Create a MathJSON solution to the following:», which then outputs a JSON between and tags, which can be parsed and passed to a deterministic library to perform calculations. Log information extraction (e.g. getting the list of paths/IPs/etc. from apache logs) Anonymization, e.g. removing names, IP addresses, and/or dates from text. Chat introspection — multi-turn chats that have random questions injected asking about the character’s descriptions, the setting of the chat, or things previously said, to make sure the model uses the system prompt and or chat history. Multi-step instructions with…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: LoneStriker
Теги: mistral, conversational, text-generation-inference, endpoints_compatible
Лайков: 2 | Загрузок: 7
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.