cyboghostginx/gemma-4-E4B-it-Adetayo-IS-non-reas
The cyboghostginx/gemma-4-E4B-it-Adetayo-IS-non-reas model is an approximately 7.9 billion parameter Gemma-4-E4B-it fine-tune, specifically specialized for Icelandic noun-phrase inflection and grammar. This model offers a stronger 4B-class Icelandic entry compared to the Gemma-3-4B fine-tune. It achieves high scores in custom 0-shot Icelandic inflection tasks, making it particularly effective for morphological accuracy in Icelandic language processing. The model is designed for direct answers, performing best when run with thinking disabled.
Loading preview...
Model Overview
The cyboghostginx/gemma-4-E4B-it-Adetayo-IS-non-reas is an Icelandic fine-tune of Google's Gemma-4-E4B-it base model, featuring approximately 7.9 billion parameters. This variant is specifically engineered to excel in Icelandic noun-phrase inflection and grammar, providing a robust solution for Icelandic language tasks in its parameter class.
Key Capabilities & Performance
- Icelandic Morphology: Achieves a notable 98.4 score on custom 0-shot inflection tasks using BÍN morphology, indicating high accuracy in handling complex Icelandic grammatical structures.
- Direct Answering: Optimized for direct responses, performing best when its 'thinking' mechanism is disabled. Users should ensure
enable_thinking=Falseis used with the chat template for optimal results. - Improved Icelandic Performance: Offers a stronger performance for Icelandic language processing compared to the Gemma-3-4B fine-tune.
Training Methodology
The model was trained using assistant-only loss masking on recombined, decontaminated BÍN data. This masking technique allows the reasoning base to effectively absorb morphology without compromising coherent chat capabilities.
Usage Recommendations
This model is ideal for applications requiring precise Icelandic noun-phrase inflection and grammatical accuracy. It can be deployed using transformers or vLLM, utilizing the model's specific chat template. Due to its specialization, it is best suited for tasks where direct, morphologically accurate Icelandic output is paramount.