Shulou(Shulou.com)11/24 Report--

Thanks to CTOnews.com netizens South China Daniel Wu and Xiao Zhan for the clue delivery! CTOnews.com, June 10, 2023 Beijing Zhiyuan artificial Intelligence Research Institute released the Wudao 3.0 model series, including the Wudao Sky Eagle (Aquila) language big model series, the "FlagEval" big language evaluation system and open platform, and the Wudao visual big model series.

CTOnews.com with details:

The Wudao Sky Eagle Aquila language model is the first bilingual model in both Chinese and English, which supports commercial use and meets the requirements of data compliance. Training starts from scratch on the basis of a high-quality Chinese-English compliance corpus database. This time, the 7B, 33B and AquilaChat dialogue models in this series of models are released, that is, the basic model of 7 billion parameters and 33 billion parameters, and the AquilaCode text code generation model.

7B and 33B technically inherit the architecture design advantages of GPT-3 and LLaMA, replace a group of more efficient underlying operators, redesign and implement Chinese-English bilingual tokenizer, upgrade the parallel training method of BMTrain, and achieve nearly 8 times training efficiency than Magtron+DeepSpeed ZeRO-2 in the training process of Aquila.

The AquilaCode-7B code model, which is based on the Aquila-7B basic pattern, is also 7B (7 billion parameters). It uses small data sets and small parameters to achieve high performance and supports both Chinese and English.

The large language model evaluation system Libra has established an omni-directional evaluation system in the three dimensions of ability, task and index, including more than 30 abilities, plus five tasks, and then multiplied by four categories of indicators, a total of almost 600 dimensions. At present, Libra open source large model evaluation system has been open to the public for registration. In terms of hardware, it supports a variety of chip architectures such as Nvidia, Huawei's Teng Teng, Cambrian and Kunlun core, as well as a variety of deep learning frameworks such as PyTorch.

In terms of visual macromodel, Zhiyuan Conference announced the achievements of Emu, the strongest billion-level visual basic model, EVA-CLIP, the most powerful open source CLIP model, Painter, the first general visual model for context image learning technology path, the universal horizon segmentation model for segmenting everything, and vid2vid-zero, the first zero-sample video editing method.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.