Jan 14, 2022|
JD Explore Academy’s Vega v1 Sets New World Record for GLUE Benchmark
by Yiming Yan
The Vega v1 model proposed by the JDExplore Dream Team beats Microsoft, Facebook, and Stanford University to top the General Language Understanding Evaluation (GLUE) list with an overall average score of 91.3, setting a new world record. GLUE is an important metric to measure natural language processing (NLP) and has been widely adopted across the industry.
Utilizing the paradigm of pre-training and fine-tuning, Vega v1 achieves breakthroughs by using several pre-training innovations such as the efficient parallelized training framework, the innovative model architecture with billions of parameters, and the improved self-supervised signals. The future Vega versions will also integrate trusted artificial intelligence and other technologies to enhance both stability and privacy while constantly improving text understanding capabilities.
The Vega v1 ranked first in four of the nine subtasks, particularly in sentiment analysis tasks The Stanford Sentiment Treebank (SST-2) and Winograd NLI (WNLI), where Vega v1 exceeded the human mind for the first time. The Academy is going to further push forward the research of smart cities, supply chain management, smart retail, and other applications, which will boost the blooming of the digital intelligence industry.
The JDExplore Dream Team is made up of the University of Sydney, Wuhan University, and Beijing University of Aeronautics and Astronautics, taking parts in tests of the GLUE Benchmark consisting of 9 tasks such as question natural language inference (QNLI), quora question pairs (QQP), and named entity recognition (NER).