OpenLLM-Ro/RoGemma2-9b-Instruct-DPO-2025-04-23
OpenLLM-Ro/RoGemma2-9b-Instruct-DPO-2025-04-23 is a 9 billion parameter instruction-tuned generative text model developed by OpenLLM-Ro, specifically designed for the Romanian language. This model is part of the RoGemma2 family, fine-tuned using Direct Preference Optimization (DPO) on various Romanian datasets, and features a 16384 token context length. It is optimized for assistant-like chat in Romanian, demonstrating strong performance across several Romanian-specific benchmarks.
Loading preview...
Model Overview
OpenLLM-Ro/RoGemma2-9b-Instruct-DPO-2025-04-23 is a 9 billion parameter instruction-tuned generative text model developed by OpenLLM-Ro, part of the RoGemma2 family. This model is specialized for the Romanian language, representing a significant open-source effort to build LLMs for Romanian. It was fine-tuned from RoGemma2-9b-Instruct-2025-04-23 using Direct Preference Optimization (DPO) on a collection of Romanian datasets including RoHelpSteer, RoUltraFeedback, RoMagpieDPO, RoArgillaMagpie, and RoHelpSteer2.
Key Capabilities
- Romanian Language Specialization: Designed and optimized specifically for generating and understanding Romanian text.
- Instruction Following: Fine-tuned for assistant-like chat and instruction-based tasks.
- Strong Benchmark Performance: Achieves an average score of 59.79 on academic benchmarks, with notable results in Winogrande (73.16), Hellaswag (64.26), and TruthfulQA (63.86). It also shows competitive performance on MT-Bench (7.26 average) and RoCulturaBench (5.36 average), consistently answering in Romanian.
Intended Use Cases
- Research in Romanian NLP: Ideal for academic and research purposes focused on the Romanian language.
- Assistant-like Chatbots: Suitable for developing conversational AI applications that interact in Romanian.
- Natural Language Tasks: Can be adapted for various Romanian natural language processing tasks, particularly those requiring instruction following.