Set a new world record! The Inspur Hairuo industry large model ranked first in the QASC international list

Recently, the global authoritative competition - QASC Challenge updated the international ranking results. The Inspur Hairuo industry large model (Hairuo), with its excellent performance and excellent reasoning ability and an absolute advantage of 93.70% accuracy, surpassed the second place of 93.48% and the average human performance of 93.33%, set a new world record, and successfully won the first place on the list! It fully proves that the Inspur Hairuo industry large model has reached the world’s leading level in language understanding and logical reasoning!

The QASC Challenge is an authoritative competition launched by the Allen Institute, the world’s top research institution, which is not only an important yardstick for evaluating the ability of machines in the field of common sense reasoning and reading comprehension, but also an important platform recognized by the industry to measure the intelligence level of natural language processing models. The evaluation requires that the participating models can accurately understand the deep meaning behind the questions, and reason based on a wide range of common sense and knowledge to finally give a reasonable answer. This challenge puts forward extremely high requirements for the language comprehension ability, knowledge integration ability and logical reasoning ability of the model, and the evaluation results are often the focus of attention in academia and industry, leading the development direction of intelligent technology.

In this challenge, with its advanced architecture and algorithms, the Hairuo industry large model comprehensively innovates and optimizes the RAG (Retrieval-Augmented Generation) technology to improve the retrieval and application ability of the machine reading comprehension system, break the limitation of relying only on the internal knowledge of the model in the traditional natural language processing task, accurately retrieve relevant knowledge from the massive knowledge base of tens of millions of levels and carry out enhanced fine-tuning training, thus improving the quality, accuracy and practicability of the generated text. At the same time, by adopting the end-to-end training strategy, and optimizing the retrieval and generating modules, the cyclic unsupervised pre-training and supervised fine-tuning training were carried out, and the strategies such as online learning were adopted to effectively use gradient information, continuously iterate the model, improve the training efficiency, take into account the certainty and generative ability of the model, and improve the accuracy of the generated model.

If 2023 is the first year of the general model, 2024 is the first year of the implementation of the industry model. Positioned as an industry large model, the first batch of the Inspur Hairuo industry large model is oriented to the six major industries of government, transportation, emergency response, manufacturing, medical care and agriculture, with three core advantages: reliable, easy to implement and sustainable. Through the adoption of MoE (hybrid expert model) architecture, comprehensive optimization of RAG, full-link security protection and other measures, key technological innovation is achieved, and operation cost, accuracy, security and other issues that industry users are concerned about in the process of large model landing are solved. Relying on Hairuo’s complete product matrix, it helps users build industry-specific large models, and based on the distributed computing power platform deployed across China, it realizes the entry with computing power, and provides more accurate and intelligent support and services for industry users.

This achievement is another breakthrough in the intelligence level of the Hairuo industry large model. In the future, Inspur Cloud will continue to increase investment in the field of large models, promote the application of Hairuo industry large models in more scenarios, and provide key support for the digital and intelligent transformation of thousands of industries. At the same time, Inspur Cloud will also work with ecological partners to embrace the wave of large models, cultivate new business formats in the industry, and promote the high-quality development of the digital economy.