iFlytek Spark V3.5 is officially released, based on the national computing power platform "Flying Star One" training

iFlytek will hold a V3.5 upgrade conference for the Spark Cognitive Large Model on January 30. Liu Qingfeng, Chairman of iFlytek, and Liu Cong, Dean of the Research Institute, officially released iFlytek Spark V3.5 based on the first national industrial computing power training.

iFlytek announced that on October 24, 2023, it would release the first Wanka domestic computing power platform "Flying Star One" that supports the training of large models with trillions of parameters, and it would be officially launched. In the more than 90 days since its launch, iFlytek Spark has launched large-scale model training with larger parameters to benchmark GPT-4 based on "Flying Star One", resulting in the iFlytek Spark V3 on January 30. 5 upgrade released.

Based on the National Open Large Model training, the National Open Large Model Fire V3.5 has achieved comprehensive results in seven aspects: language understanding, text generation, knowledge question and answer, logical reasoning, mathematical ability, coding ability and multi-modal ability. upgrade. Among them, the language understanding and mathematical ability exceed GPT-4 Turbo, the code reaches 96% of GPT-4 Turbo, and the multi-modal understanding reaches 91% of GPT-4V.

Currently, iFlytek Spark has empowered leading enterprises in the fields of insurance, banking, energy, automobiles, communications, etc., through cooperation with Pacific Insurance, through cooperation with Spark Pacific Insurance, and through the Spark Pacific Plan to empower internal knowledge services , office, auditing, exhibition industry, etc., to create a benchmark for the application of digital labor in the insurance field; jointly with the Bank of Communications, empowering customer service, exhibition industry, office, research and development, etc., focusing on creating a benchmark for the application of code capabilities in the banking field; jointly with the National Energy Group, empowering It can create a large-scale model application benchmark for the integrated linkage of central enterprise groups in coal, electric power, transportation, chemical industry, etc.; it has joined forces with the National Energy Group to empower coal, electric power, transportation, chemical industry, etc., and create a large-scale model of integrated linkage of central enterprise groups. Application benchmarks.

In addition, the "iFlytek Spark Open Source Large Model", which is deeply adapted to domestic computing power, was released for the first time, with leading scene application effects, and the Shengsi open source community jointly launched it online.

This open source has 130 million parameters (13B), including the basic model iFlytekSpark-13B-base, the fine-tuning model iFlytekSpark-13B-chat, and the fine-tuning tool iFlytekSpark- 13B-Lora, personality customization tool iFlytekSpark-13B-Charater. Academic enterprise research can more conveniently train their own dedicated large models based on the full-stack autonomous and controllable Spark optimization suite.

Liu Qingfeng revealed that the Spark open source large model has formed a differentiated advantage in technology. Spark Open Source-13B ranks among the best in a number of well-known public evaluation tasks. In typical enterprise scenarios such as text generation, language understanding, text rewriting, industry Q&A, machine translation, etc., through in-depth research and optimization in areas such as learning assistance and language understanding, The performance has been greatly improved and it is more convenient when processing complex natural language tasks.

Based on the "Flying Star One" training, the full-stack domestic adaptation and optimization of the Spark open source large model is simple and easy to use, with leading scene application effects. The training strategy is extremely optimized for Ascend computing power, and the training efficiency reaches A100 90%. This is not only a further in-depth optimization of Ascend AI hardware, but also demonstrates the determination and ability of domestic computing power to catch up with international advanced levels.

Open source can better enhance ecological cooperation only if more scenarios are implemented. At present, Huawei Shengsi Open Source Community has officially launched the Spark Large Model Open Source Version-13B, which is completely free for academic and corporate research, enhancing academic cooperation and industrial exploration at the same time.

Looking forward to the development of the iFlytek Spark large model in 2024, Liu Qingfeng pointed out three points: "First, we must continue to benchmark the underlying capabilities of the general large model against the most advanced international levels, from algorithm research to including smaller calculations. We must be clear-headed and rational to see the gap. Currently, there is still a gap between the best level of GPT-4 in areas such as small sample rapid training, multi-modal deep learning training, and ultra-complex deep understanding. iFlytek Spark is confident that it can catch up with the current best level of GPT-4 in the first half of this year. “Today’s general large models do not necessarily represent the entire future of artificial intelligence. There is still a lot of innovation to be done, such as brain science interaction, deep connection of adversarial networks, etc., which require the entire innovative ecosystem, but we must have the courage and expectations to go forward. At the forefront." Liu Qingfeng pointed out.

"Second, in 2024 we need to truly make the quality and quantity of large models take off, not only in terms of industry applications, but also in linking large models in many key technological innovations. The Chinese business community and the scientific community are confident that they can achieve excellence. "

"Third, high-rise buildings must be built on safe and controllable platforms, and we must achieve ecological prosperity on independent and controllable platforms." Liu Qingfeng said that he is confident that he can achieve the goal from algorithms and data. , application scenarios to computing power, to build a prosperous artificial intelligence ecosystem that is completely autonomous and controllable.

The above is the detailed content of iFlytek Spark V3.5 is officially released, based on the national computing power platform "Flying Star One" training. For more information, please follow other related articles on the PHP Chinese website!