Overview
GaMS-Beta/GaMS-9B-SFT-Translator-DPO is a 9 billion parameter language model developed by a team of researchers at the University of Ljubljana, Faculty for Computer and Information Science. It is a fine-tuned version of GaMS-9B-SFT-Translator, enhanced through Direct Preference Optimization (DPO). The model's training data was synthetically generated by using GaMS-9B-SFT-Translator and EuroLLM-9B-Instruct, with translations ranked by automatic metrics.
Key Capabilities
- Multilingual Machine Translation: Primarily supports Slovene and English, with additional capabilities in Croatian, Bosnian, and Serbian. It may also function for other languages supported by its Gemma 2 base.
- DPO Fine-tuning: Utilizes Direct Preference Optimization for improved translation quality, based on a preference dataset derived from multiple translation models.
- Base Model: Built upon the cjvt/GaMS-9B-Instruct architecture.
Evaluation and Performance
The model was evaluated using a custom script across various data types (ccnews, nemotron, wikipedia). While its "Overall Comet" score (0.708042) is slightly lower than gemini-2.5-flash (0.717982) and GaMS-9B-Instruct-DPO-Translator (0.714729), it demonstrates competitive performance in translation quality. Notably, it exhibits a low percentage of "Bad Lang" (0.91%) and "Short" (0.28%) translations, though it has a higher "Bad Markdown" percentage (18.28%) compared to some counterparts.
Good for
- Slovene-English Translation: Ideal for applications requiring high-quality translation between Slovene and English.
- Balkan Language Translation: Suitable for use cases involving Croatian, Bosnian, and Serbian translation.
- Research in DPO for Translation: Provides a strong baseline for further research into applying DPO to machine translation tasks, particularly with synthetic preference data.