Kuaishou’s Self-Developed Large-Scale Model ‘KwaiYii’ Makes Its Debut

Recently, Kuaishou’s self-developed giant language mannequin ‘KwaiYii‘ has entered inner testing and offered normal APIs and customised venture collaboration options for the enterprise staff.

In the CMMLU Chinese language-oriented basis mannequin record, the spectacular 13B model KwaiYii-13B ranks first in each five-shot and zero-shot classes. It demonstrates sturdy efficiency in humanities, particular Chinese matters, and achieves a mean rating of over 61 factors.

Upon looking the GitHub web page, it was discovered that the official description states: KwaiYii, developed independently by the Kuaishou AI staff, is a sequence of large-scale language fashions (LLM) constructed from scratch. Currently, it contains fashions with varied parameter sizes, such because the pre-training mannequin (KwaiYii-Base) and the chat mannequin (KwaiYii-Chat). Here we introduce the KwaiYii-13B sequence mannequin, which has a scale of 13 billion parameters.

Its foremost options embrace: KwaiYii-13B-Base pre-trained mannequin has wonderful basic technical capabilities and achieves state-of-the-art efficiency in most authoritative Chinese/English benchmarks with the identical mannequin measurement. For instance, the KwaiYii-13B-Base pre-trained mannequin is presently main in benchmarks resembling MMLU, CMMLU, C-Eval, HumanEval on the identical mannequin scale.

SEE ALSO: Kuaishou Incentivizes User Collaboration with 60 Billion Network Traffic

KwaiYii-13B-Chat dialogue mannequin has wonderful language understanding and technology capabilities, supporting a variety of duties resembling content material creation, info session, mathematical logic, code writing, and multi-turn conversations. The outcomes of guide analysis present that KwaiYii-13B-Chat surpasses mainstream open-source fashions and approaches the extent of ChatGPT (3.5) in content material creation, info session, and mathematical problem-solving.

According to reviews, the AI staff at Kuaishou will proceed to iterate on the ‘KwaiYii’ giant mannequin. On the one hand, they may proceed to optimize mannequin efficiency and develop multimodal capabilities. On the opposite hand, they’re additionally selling implementation in additional C-end and B-end enterprise eventualities.

Sign up at present for five free articles month-to-month!