OPPO announced that its artificial intelligence large model, AndesGPT-2.0, has excelled in the authoritative third-party evaluation agency SuperCLUE's "Chinese Large Model Benchmark Evaluation Report for August 2024," winning multiple first-place awards, signifying further industry recognition of OPPO's technological prowess in AI.
AndesGPT-2.0 ranks second among domestic large models on the overall leaderboard and has secured the top spot in three sub-categories: SuperCLUE-Safety, SuperCLUE-Math6, and SuperCLUE-Agent. The model has achieved significant breakthroughs in key technologies such as dialogue enhancement, efficient reinforcement learning, and multimodal capabilities, demonstrating its industry-leading positions in AI technology safety, mathematical reasoning, and agent capabilities.
As the core engine of OPPO's AI strategy, AndesGPT-2.0 is fully empowering OPPO's smart terminals, supporting diverse application scenarios including call assistants and intelligent summaries. The company plans to deploy models of various levels through its self-built OPPO AI Binhai Bay Data Center to address various application scenarios, achieving efficient deployment through end-cloud collaboration.
The three core advantages of AndesGPT-2.0 include dialogue enhancement technology, efficient reinforcement learning strategies, and multimodal capabilities. In dialogue enhancement, the model has significantly improved the user experience by deeply researching multi-turn dialogues and adopting multi-turn instruction data and multi-turn preference data synthesis techniques. In training methods, AndesGPT-2.0 has improved the intelligence and reliability of the model through iterative supervised fine-tuning and preference alignment optimization methods. Additionally, AndesGPT-2.0 has upgraded the large language model to a multimodal large model, transforming the interaction method from traditional language interaction to language + visual + touch multimodal interaction, providing a more natural and smooth interaction experience.
In terms of technological breakthroughs, OPPO's R&D team has comprehensively enhanced the overall performance of AndesGPT-2.0 by meticulously collecting and optimizing instruction datasets. In terms of safety, OPPO has established an AI security lab, focusing on tracking the latest attack methods in the industry and comprehensively evaluating the safety of its self-developed models. In terms of agent capabilities, AndesGPT has achieved efficient execution of automated task flows such as quick commands by constructing a "tool retrieval + task planning" model. In terms of mathematical reasoning capabilities, the AndesGPT R&D team has collected a vast amount of open-source mathematical datasets and significantly enhanced the model's mathematical abilities through triple data filtering and constructing high-quality data.