A major step has been taken in the field of Turkish Natural Language Processing (NLP). VBART has emerged as the first large language (artificial intelligence) model trained from scratch for Turkish. Developed by VNGRS, VBART is inspired by the BART and mBART models and is presented in two sizes: Large and XLarge.
VBART stands out as the first large-scale language model (artificial intelligence) designed by Turkish engineers and trained from scratch for Turkish. This model aims to provide accurate and context-aware meaning for a wide range of Turkish LLM applications. VBART models outperform even the latest AI products in tasks such as text summarization, heading generation, text acronymization, question answering, and question generation. These models perform three times better than multilingual models, and the monolingual tokenizer, specifically trained for Turkish, is 11 times more efficient than multilingual tokenizers.
VBART is paving a new path in Turkish NLP research by offering the possibility of fine-tuning for future text generation tasks and datasets.
VBART models were trained on 135 GB of purified Turkish data through 2.7 million steps and exposed to 708 billion tokens. This process enabled the models to learn the Turkish language in depth and achieve high accuracy and context awareness in text generation.
VBART models, tokenizers, and the cleaned vngrs-web-corpus have been made available to researchers and developers. This is an important step that will enable further research and development in the field of Turkish NLP.
VBART stands out as a revolutionary development in the field of Turkish Natural Language Processing. Its superior performance and efficiency, the opportunities it offers for future research, and its publicly available resources are opening a new era in Turkish NLP research.
VBART models and related resources are available on the Hugging Face platform. You can find more information at the following links:
