Dorna2-Llama3.1-8B-Instruct Overview

Dorna2-Llama3.1-8B-Instruct is an 8 billion parameter, decoder-only instruction-tuned model developed by Part AI. It is built on the Meta Llama 3.1 Instruct architecture and has been specifically fine-tuned using extensive Persian datasets. This specialization makes it particularly adept at handling Persian language tasks and understanding cultural nuances.

Key Capabilities & Performance

The model's primary strength lies in its proficiency with the Persian language. It has been evaluated across five specialized Persian tasks, including:

• Part Multiple Choice: Common knowledge and reasoning.
• ARC Easy & ARC Challenge: General knowledge and advanced reasoning.
• MMLU Pro: Professional-level examinations.
• AUT Multiple Choice Persian: Specialized Persian-language examination.

In comparative evaluations against other Llama 3.1-8B models, Dorna2-Llama3.1-8B-Instruct achieved an average accuracy of 50.72% across these Persian benchmarks. Notably, it scored highest in ARC Easy (79.59%), ARC Challenge (64.42%), and MMLU Pro (21.47%) among the compared models, demonstrating strong reasoning and knowledge recall in Persian contexts. The evaluation datasets comprise over 40,000 samples, ensuring a robust testing ground for Persian LLMs.

Ideal Use Cases

• Persian Language Applications: Excellent for chatbots, content generation, and translation services targeting Persian speakers.
• Research & Development: Suitable for researchers working on LLMs with a focus on low-resource or non-English languages, particularly Persian.
• Educational Tools: Can be integrated into educational platforms requiring advanced Persian language comprehension and generation.