This model is an important and more balanced improvement over our previous model. To get early access to this model you need to join to the CEAMFA community by keeping an active ko-fi subscription. You can get more information on how to 👉access this model by clicking here. You can also wait for the Open-Weight Release Date if you don’t want to pay for a subscription. — Temp 0 for best results. — GGUF first, once everything is confirmed to be working out of the box, safetensors goes next. — No issues anymore. — Doom looping and language switching are core recurrent issues — The model has a known preference to use an unused token. We are still investigating if it is happening due to a mismatch in the tokenizer during the knowledge distillation phase or an issue with gguf, we are working on a fix soon. — It looks like a tokenization issue, after more tests and token switching done the model improved on most benchmarks even more while degrading arc_challenge and significantly degrading ifeval — Currently targetting reviews on specific layers from the model to understand the issue and try to heal it from its distillation «state» — Indeed was a tokenization issue, fix and updates are…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: CEAMFA
Теги: gguf, llama-cpp, en, endpoints_compatible, conversational
Лайков: 4 | Загрузок: 4
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.