Base model: google/gemma-4-26b-it Architecture: MoE — 26B total / ≈4B active parameters (1 shared expert + 8 routed from a pool of 128 per MoE layer, 30 MoE layers) Method: Activation-directed expert surgery — 128 → 64 experts per layer (50% reduction) Quantization: Q4KM (≈9.7 GB on disk) Tags: JThomas-CoE/coe-gemma4-[domain]-mmlupro-14b-a4b-q4 | Ollama: coe-gemma4-[domain]-mmlupro-14b-a4b:q4 Hub: JThomas-CoE on HuggingFace This README covers the MMLU-Pro–derived domain speci The key claim this release tests is that domain-relevant expert populations can be identified using existing public benchmark questions alone, making the surgery pipeline fully automatable. All models in this series must be used in thinking-off mode. If you are using the Ollama API, pass «think»: false in your request body. If you are accessing the model via a raw API (llama.cpp server, OpenAI-compatible endpoint, etc.) you must inject a closed thinking block at the start of the assistant turn: Why this is required: Expert surgery retains 50% of…
Модальности:
Генерация текста
Области применения:
Математика Биология Химия
Задача: Генерация текста
Автор: JThomas-CoE
Теги: gguf, moe, mixture-of-experts, domain-specialist, expert-surgery, mmlu-pro, college-of-experts, ollama
Лайков: 3 | Загрузок: 46
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
