cjvt/GaMS2-27B-GRPO-Translator
The cjvt/GaMS2-27B-GRPO-Translator is a 27 billion parameter language model developed by the University of Ljubljana, Faculty for Computer and Information Science, specifically fine-tuned for high-quality English to Slovene translation. Built upon the GaMS2-27B base model, it underwent a three-phase training process including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Generative Reward Policy Optimization (GRPO) to ensure natural, fluent, and accurate translations. This model excels at contextual and idiomatic translation, avoiding literal interpretations, and is designed for professional translation tasks.
Loading preview...
GaMS2-27B-GRPO-Translator: Specialized English to Slovene Translation
GaMS2-27B-GRPO-Translator is a 27 billion parameter model developed by the University of Ljubljana, Faculty for Computer and Information Science, specifically engineered for high-quality English to Slovene translation. It is built on the cjvt/GaMS2-27B base model and licensed under Gemma terms.
Key Capabilities
- Professional Translation: Optimized to produce natural, fluent, and accurate translations from English to Slovene.
- Contextual Understanding: Avoids literal translations by considering context, idioms, and cultural references.
- Refined Training: Underwent a three-phase fine-tuning process:
- SFT: Supervised Fine-Tuning on 1000 high-quality examples to establish proper chat template usage.
- DPO: Direct Preference Optimization on 1500 examples to correct formatting errors, truncation, and premature stopping.
- GRPO: Generative Reward Policy Optimization on 100k examples using various reward functions (COMET, length, language scores) for main training.
- Standard Chat/Instruct Interface: Compatible with standard chat templates, similar to cjvt/GaMS2-27B-Instruct.
Good For
- High-Quality English to Slovene Translation: Ideal for applications requiring precise and natural language translation.
- Developers: Can be easily integrated using the
transformersorvLLMlibraries, with provided code examples. - Research: Provides a specialized model for studying advanced translation techniques, particularly GRPO.