Metin/Gemma-2-9b-it-TR-DPO-V1 is a 9 billion parameter instruction-tuned Gemma-2 model, fine-tuned by Metin Usta using Direct Preference Optimization (DPO). This model is specifically optimized for enhancing output format and content quality in Turkish, making its responses more fluent, coherent, and informative compared to its base model. It was trained on a synthetically generated preference dataset of 10K samples to improve Turkish language generation.
Loading preview...
Overview
Metin/Gemma-2-9b-it-TR-DPO-V1 is a 9 billion parameter instruction-tuned model, derived from Google's gemma-2-9b-it base model. Developed by Metin Usta, this version has undergone Direct Preference Optimization (DPO) using a synthetically generated preference dataset of 10,000 samples. The primary goal of this fine-tuning was to enhance the output format and content quality specifically for the Turkish language.
Key Capabilities
- Improved Turkish Fluency: Generates more fluent and coherent responses in Turkish compared to the base model.
- Enhanced Content Quality: Produces more informative and detailed answers for given instructions in Turkish.
- Preference-Tuned: Optimized through DPO to yield more "likable" and preferable outputs.
Performance
The model was trained for 2 hours on a single NVIDIA H100 GPU. Benchmarks on the OpenLLMTurkishLeaderboard_v0.2 show an average score of 55.13%, with specific scores including 51.69% on MMLU_TR_V0.2 and 65.07% on GSM8K_TR_V0.2.
Use Cases
This model is particularly well-suited for applications requiring high-quality, coherent, and informative text generation in Turkish. While it aims for improved output preference, users should still verify outputs as the model may occasionally generate incorrect or nonsensical content.