OrisTeam/Koliber-v1.1-Preview - Каталог нейросетей
Генерация текста

OrisTeam/Koliber-v1.1-Preview

Добавлено:
OrisTeam/Koliber-v1.1-Preview

Koliber-v1.1-Preview is an experimental continued-training checkpoint of Koliber v1.0 Base, a compact Polish decoder-only causal language model developed by OrisTeam. It remains a 126M base pretrained language model, not an instruction-tuned assistant or chatbot. The goal of v1.1 was deliberately narrow: improve local next-token behavior and reduce some undesirable generation drift without increasing model size, changing the tokenizer, or replacing the underlying Koliber architecture. > [!WARNING] > Koliber v1.1-Preview is an unaligned base language model. > > Generated text may be false, misleading, offensive, biased, unsafe, inconsistent, or otherwise inappropriate. The model may invent facts, names, quotations, dates, numbers, sources, or events, and may reproduce undesirable patterns present in its training data. > > Outputs should not be treated as factual, authoritative, or safe by default. Additional evaluation, filtering, alignment, and application-specific safeguards are required before deployment in user-facing or high-stakes settings. Koliber v1.1 does not introduce a larger architecture or a new tokenizer. The released inference model keeps the same core model shape as…

Модальности:
Генерация текста


Задача: Генерация текста
Автор: OrisTeam
Теги: koliber, polish, causal-lm, base-model, continued-pretraining, teacher-guided, gqa, rope
Лайков: 3  |  Загрузок: 115

Открыть на HuggingFace →

Описание основано на материалах HuggingFace. Перевод выполнен автоматически.