Sensitivity-guided mixed-precision quantization of coder3101/gemma-4-E4B-it-heretic, which is itself a decensored variant of google/gemma-4-E4B-it produced by coder3101 using Heretic v1.2.0. All future Heretic versions of this build will live in this repository. Version identifiers appear only in filenames, not in the repo name. No vision mmproj is included. Stock Cerebellum v2 ships none either — Gemma 4 E4B’s vision projector is not publicly available in a distributable form. 1. Base architecture: google/gemma-4-E4B-it — Google DeepMind (Apache-2.0) 2. Heretic variant: coder3101/gemma-4-E4B-it-heretic — coder3101. The safetensors from that repository were converted to F16 GGUF and used as the direct quantization source (single quantization step from F16). coder3101 applied Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation), targeting layers 21–42. Their reported result: 0.0058 KL divergence from base, 3/100 refusals vs 99/100 on the original model. 3. Quantization: Cerebellum v2 recipe transferred verbatim from the stock deucebucket/Gemma-4-E4B-it-Cerebellum-v2-GGUF build — same 179-entry tensor-type override file, same imatrix. The heretic…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: deucebucket
Теги: gguf, GGUF, gemma4, gemma, quantized, cerebellum, imatrix, mixed-precision
Лайков: 3 | Загрузок: 221
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
