AlicanKiraz0/Kizagan-E4B-Turkish-Reasoning-Model
AlicanKiraz0/Kizagan-E4B-Turkish-Reasoning-Model is a 7.9 billion parameter, instruction-tuned causal language model developed by Alican Kiraz. Fine-tuned from Google's Gemma 4 E4B-it, it specializes in Turkish language understanding and mathematical reasoning, showing significant improvements on Turkish MMLU and GSM8K benchmarks. This model is optimized for consistent output across diverse Turkish use cases, from daily dialogue to technical and academic scenarios.
Loading preview...
Model Overview
Kızagan-E4B is a 7.9 billion parameter Turkish reasoning model developed by Alican Kiraz, fine-tuned from Google DeepMind's Gemma 4 E4B-it. It leverages a 540,000-line (~2 billion token) Turkish SFT dataset, meticulously designed with a multi-dimensional taxonomy to enhance its understanding of Turkish linguistic nuances and reasoning capabilities.
Key Capabilities
- Enhanced Turkish Reasoning: Achieves a +21.88 point increase on Turkish GSM8K and +2.76% on Turkish MMLU compared to its base model, demonstrating strong mathematical and general knowledge improvements.
- Comprehensive Taxonomy: Trained on a dataset structured across 8 domains, 7 skill families, 25+ sub-skills, and 9 personas, ensuring robust and consistent output for various scenarios.
- Multi-format Support: Available in multiple variants including Transformers (BF16), GGUF (for
llama.cpp, LM Studio, Ollama), and MLX (for Apple Silicon devices).
Good For
- Multi-step instruction following: Generating outputs with specific formats and constraints.
- Turkish text manipulation: Summarization, formalization, simplification, and expansion of Turkish content.
- Structured extraction: Converting free-form text into JSON or tabular formats.
- Grounded Q&A: Providing evidence-based answers from given passages, reducing hallucinations.
- Mathematical and everyday reasoning: Solving math word problems, performing error analysis, and supporting decision-making.
- Turkish linguistic tasks: Handling orthography, punctuation, morphology, and noisy text normalization.