cyboghostginx/gemma-4-31B-it-Adetayo
The cyboghostginx/gemma-4-31B-it-Adetayo model is an Icelandic fine-tune of Google's Gemma 4 31B instruction-tuned large language model, featuring 31 billion parameters and a 32768-token context length. This model is specifically optimized for Icelandic morphology and grammar, demonstrating improved performance in these areas compared to its base model when evaluated without a reasoning path. It is designed for standard text generation tasks in Icelandic, leveraging the Gemma chat template for interaction.
Loading preview...
Model Overview
cyboghostginx/gemma-4-31B-it-Adetayo is an Icelandic fine-tune of Google's Gemma 4 31B instruction-tuned model. This 31 billion parameter model is specifically adapted to target Icelandic morphology and grammar, building upon the base Gemma 4 architecture. It maintains a substantial context length of 32768 tokens, making it suitable for processing longer Icelandic texts.
Key Capabilities
- Icelandic Language Optimization: The model shows gains in Icelandic morphology compared to the base Gemma 4 31B-it model, particularly when evaluated without a reasoning path.
- Instruction-Tuned: Designed for instruction-following, it utilizes the standard Gemma chat template for interaction.
- Standard Text Generation: Primarily intended for general text generation tasks in Icelandic.
Performance Notes
On the Miðeind Icelandic LLM leaderboard, this model achieved a 6-task average score of 70.09. It's important to note that this evaluation was conducted with the model's "thinking" path disabled, unlike the base Gemma-4-31B-it which uses its reasoning path. A local 5-task average (thinking off) for this fine-tune reached 83.59, indicating its strong performance in direct generation tasks. The fine-tuning focused on improving inflection, which was the primary area of gain.
Usage
Users should apply the Gemma chat template (tokenizer.apply_chat_template) for interaction and load the model using Gemma4ForConditionalGeneration on a GPU. Evaluation should be performed with thinking disabled to align with its optimized performance profile.