GaMS-Beta/GaMS-9B-Instruct-DPO-Translator
GaMS-9B-Instruct-DPO-Translator is a 9 billion parameter instruction-tuned causal language model developed by researchers at the University of Ljubljana, Faculty for Computer and Information Science. This model is a fine-tuned version of GaMS-9B-Instruct, optimized using Direct Preference Optimization (DPO) with synthetically generated translation data. It specializes in machine translation, primarily supporting Slovene and English, with secondary support for Croatian, Bosnian, and Serbian, and a context length of 16384 tokens.
Loading preview...
Overview
GaMS-9B-Instruct-DPO-Translator is a 9 billion parameter instruction-tuned model developed by a research team at the University of Ljubljana. It is a DPO-fine-tuned variant of the GaMS-9B-Instruct base model, specifically designed for machine translation tasks. The model's training leveraged synthetic data generated by GaMS-9B-SFT-Translator and EuroLLM-9B-Instruct, with preference ranking based on automatic translation quality metrics.
Key Capabilities
- Multilingual 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.
- Direct Preference Optimization (DPO): Enhanced for translation quality through DPO, utilizing a dataset where translations were ranked for reliability.
- Competitive Performance: Evaluation shows it achieves a competitive Overall Comet score of 0.714729, performing closely to gemini-2.5-flash on various translation benchmarks like ccnews and wikipedia.
Good For
- Machine Translation: Ideal for translating between Slovene, English, Croatian, Bosnian, and Serbian.
- Research in DPO and Synthetic Data: Useful for researchers exploring the application of DPO with synthetically generated preference data for translation tasks. The evaluation script and data are publicly available for further study.