RoGemma2-9b-Instruct-DPO-2024-10-09 Overview
This model is a 9 billion parameter instruction-tuned generative text model developed by OpenLLM-Ro, specifically designed for the Romanian language. It represents the first open-source effort to build a large language model specialized for Romanian, offering both foundational and instruct/chat variants. This particular version is a human-aligned instruct model, fine-tuned using Direct Preference Optimization (DPO) from the RoGemma2-9b-Instruct-2024-10-09 base model, utilizing the RoHelpSteer dataset.
Key Capabilities & Performance
- Romanian Language Specialization: Optimized for various natural language tasks in Romanian, including assistant-like chat.
- Academic Benchmarks: Achieves an average score of 59.08 on academic benchmarks, outperforming its base model and gemma-2-9b-it in several categories, including a notable 57.24 on GSM8k.
- Downstream Task Performance: Demonstrates strong results in Romanian-specific downstream tasks:
- LaRoSeDa: Achieves 97.74 Macro F1 for Binary classification and 67.40 Macro F1 for Multiclass classification in few-shot settings.
- WMT (EN-RO): Scores 27.32 BLEU in few-shot English-to-Romanian translation.
- STS: Records 80.82 Spearman and 81.50 Pearson correlation in few-shot semantic textual similarity.
- MT-Bench & RoCulturaBench: Shows competitive performance in Romanian-specific conversational and cultural understanding benchmarks, with 6.77 average on MT-Bench and 4.83 on RoCulturaBench, consistently providing answers in Romanian.
Intended Use Cases
This model is primarily intended for research use in Romanian. Its instruction and chat-tuned variants are suitable for assistant-like chat applications, while the base models can be adapted for a variety of other natural language tasks within the Romanian language context. Use in any manner that violates the cc-by-nc-4.0 license or applicable regulations, or use in languages other than Romanian, is considered out-of-scope.