OpenLLM-Ro/RoLlama3-8b-Instruct-DPO-2025-04-23
OpenLLM-Ro/RoLlama3-8b-Instruct-DPO-2025-04-23 is an 8 billion parameter instruction-tuned generative text model developed by OpenLLM-Ro, built upon Meta Llama 3. This model is specifically fine-tuned for the Romanian language, excelling in assistant-like chat applications. It is distinguished by its human-aligned DPO training, making it a specialized tool for research and natural language tasks in Romanian.
Loading preview...
Model Overview
OpenLLM-Ro/RoLlama3-8b-Instruct-DPO-2025-04-23 is an 8 billion parameter instruction-tuned generative text model developed by OpenLLM-Ro, part of the RoLlama3 family built with Meta Llama 3. This specific variant is a human-aligned instruct model, fine-tuned using Direct Preference Optimization (DPO) on a collection of Romanian datasets including RoHelpSteer, RoUltraFeedback, RoMagpieDPO, RoArgillaMagpie, and RoHelpSteer2. It is designed for assistant-like chat applications and research in Romanian.
Key Capabilities and Performance
- Romanian Language Specialization: Represents the first open-source effort to build an LLM specifically for Romanian, offering both foundational and instruct/chat variants.
- Instruction Following: Optimized for assistant-like interactions in Romanian, demonstrating strong performance in conversational tasks.
- Benchmarking: Achieves an average score of 55.86 on academic benchmarks, with notable scores of 52.26 on ARC, 57.06 on TruthfulQA, and 6.67 on the Romanian MT-Bench average, outperforming other RoLlama3-8b-Instruct models in several metrics.
- Downstream Tasks: Shows strong few-shot performance on LaRoSeDa Binary (Macro F1: 97.60).
Intended Use Cases
- Research: Primarily intended for research purposes in the Romanian language.
- Assistant-like Chat: Suitable for developing conversational AI applications that require high proficiency in Romanian.
Limitations
- Use is restricted to the Romanian language; performance in other languages is not guaranteed.
- Subject to the cc-by-nc-4.0 license, prohibiting commercial use.