ytu-ce-cosmos/Turkish-Gemma-9b-T1
Hugging Face Overview
ytu-ce-cosmos/Turkish-Gemma-9b-T1 is a 9 billion parameter model based on the Gemma architecture, developed by ytu-ce-cosmos. It is an adaptation of ytu-ce-cosmos/Turkish-Gemma-9b-v0.1, specifically enhanced for multi-step reasoning in Turkish. The model focuses on improving performance in complex problem-solving scenarios.
Key Capabilities
- Multi-step Reasoning: Stronger intermediate inference for tasks involving multiple clues or conditions.
- Math & Logic: Improved accuracy across arithmetic, probability, sequences, rational reasoning, and logic puzzles.
- Instruction Following: Better adherence to user prompts.
- Reduced Hallucinations: Provides more grounded answers and indicates uncertainty when appropriate.
Performance
Evaluated against a human-annotated dataset of 1,450 questions, Turkish-Gemma-9b-T1 achieved a 68.65% win rate, outperforming several larger models like Qwen3-32B and Google's Gemma-3-27b-it. On the Turkish Gsm8k benchmark, it scored 77.41, placing it competitively among high-parameter models. The developers emphasize the importance of adjusting evaluation configurations for reasoning models to accurately reflect their performance.
Usage Tips
For optimal performance, it is recommended to use specific generation parameters: Temperature=0.6, TopP=0.95, TopK=20, and MinP=0. Greedy decoding should be avoided as it can lead to performance degradation and repetitions. For complex tasks, increasing max_new_tokens and adjusting repetition_penalty or presence_penalty can help manage output quality.