mmaccrate/gemma-4-E2B-it-QAD-GGUF - Каталог нейросетей
Генерация текста

mmaccrate/gemma-4-E2B-it-QAD-GGUF

Добавлено:
mmaccrate/gemma-4-E2B-it-QAD-GGUF

A Q4_0 GGUF of Google’s Gemma 4 E2B IT, trained with quantization-aware distillation (QAD). QAD trains a quantized student against token-level distributions from a frozen full-precision teacher. This lets the student adapt to the precision loss introduced by quantization instead of applying quantization only after training is complete. — Model file: gemma-4-e2b-it-qad-q40.gguf — Format: Q40 GGUF — Base model: google/gemma-4-E2B-it — Runtime tested: llama.cpp — Project article: QAD, Distilled — Source repository: mmaccrate/maccrate.ai The GGUF uses a mixed layout: Q40 for most layers, with Q6K for the token embedding and final vocabulary projection to preserve stability. The artifact was exported and evaluated with llama.cpp revision 95ef7fc16054e63b427a3ef00188e055ef7586d8. Newer revisions may work, but use the pinned revision for matched numerical comparisons. The frozen teacher and trainable student both began from google/gemma-4-E2B-it revision 3e22461f65e89153144f8adb70e3b8c2cc9845a7. — Frozen BF16 teacher — All student language-model parameters trainable — Mixed GGUF fake quantization active during the student forward and backward pass — Q40 linear weights with Q6K embedding…

Модальности:
Генерация текста

Области применения:
Диалог / чат


Задача: Генерация текста
Автор: mmaccrate
Теги: gguf, q4_0, quantization-aware-training, knowledge-distillation, gemma, en, endpoints_compatible, conversational
Лайков: 3  |  Загрузок: 74

Открыть на HuggingFace →

Описание основано на материалах HuggingFace. Перевод выполнен автоматически.