Overview

OpenLLM-Ro/RoLlama3-8b-Instruct is an 8 billion parameter instruction-tuned generative text model developed by OpenLLM-Ro, built upon Meta-Llama-3-8B-Instruct. This model is part of the first open-source initiative to create a comprehensive family of LLMs specifically for the Romanian language, including foundational, instruct, and chat variants. It is primarily intended for research and assistant-like chat applications within the Romanian linguistic context.

Key Capabilities & Training

This model has been fine-tuned using a diverse set of Romanian instruction-following datasets, including RoAlpaca, RoDolly, RoSelfInstruct, and RoUltraChat, among others. This extensive training on Romanian-specific data aims to optimize its performance for tasks requiring understanding and generation in Romanian.

Performance Highlights

RoLlama3-8b-Instruct demonstrates competitive performance against its base model, Llama-3-8B-Instruct, on various benchmarks. Notably, the RoLlama3-8b-Instruct-DPO-2025-04-23 variant achieves an average score of 55.86 on academic benchmarks, with strong results in MMLU (55.35), Hellaswag (59.93), and GSM8k (43.95). On the MT-Bench for Romanian, it scores an average of 6.67, and consistently answers in Romanian (160/160). It also performs well on the RoCulturaBench, scoring 4.83.

Intended Use Cases

• Research: Ideal for academic research focused on natural language processing in Romanian.
• Assistant-like Chat: Suitable for developing conversational AI applications that interact in Romanian.

Limitations

This model is specifically designed for Romanian and its use in other languages is out-of-scope. Users should also be aware of the license (cc-by-nc-4.0) and ensure compliance with all applicable laws and regulations.