Core Concepts
Peacock introduces a family of Arabic Multimodal Large Language Models, addressing the lack of resources in languages other than English and showcasing strong performance in visual reasoning tasks.
Abstract
Peacock presents a comprehensive suite of Arabic MLLMs, Peacock models outperform multilingual baseline mBlip, Pretraining data is curated from publicly available English datasets translated into Arabic, Models trained in two stages - pretraining and instruction finetuning, Peacock models excel in VQA tasks on various datasets, Henna benchmark evaluates model capabilities related to Arabic culture.
Stats
Native population of Arabic over 400 million speakers.
SEED-Bench comprises 19K multiple-choice questions.
LLaVA-Bench includes 30 images with three questions each.
Henna benchmark consists of 1,132 samples related to Arabic culture.