Gamingdeputy News on June 10th, at the “2023 Beijing Zhiyuan Conference”, Beijing Zhiyuan Artificial Intelligence Research Institute released the Enlightenment 3.0 large-scale model series, includingEnlightenment·Aquila language large-scale model series, “FlagEval” large language evaluation system and open platform, and Enlightenment·Vision visual large-scale model series.
Gamingdeputy attaches detailed content:
Enlightenment·Aquila language model is the first bilingual model in Chinese and English, supports commercial use, complies with data compliance requirements, and trains from scratch on the basis of Chinese and English high-quality and compliant corpus databases.Released this time in the series of models 7B, 33B and AquilaChat dialogue models, that is, the basic model with 7 billion parameters and 33 billion parameters, and AquilaCode text code generation model.

7B and 33B technically inherit the architectural design advantages of GPT-3, LLaMA, etc., replace a batch of more efficient underlying operator implementations, redesign and realize the tokenizer in Chinese and English, and upgrade the BMTrain parallel training method.In the training process of Aquila, the training efficiency is nearly 8 times higher than that of Magtron+DeepSpeed ZeRO-2.
The AquilaCode-7B code model is a code model based on the Aquila-7B basic model, which is also 7B (7 billion parameters),It achieves high performance with small data sets and small parameters, and supports Chinese and English bilinguals.
The large language model evaluation system “Libra” has established a comprehensive evaluation system in three dimensions: ability, task, and index, including more than 30 abilities, plus 5 types of tasks, and then multiplied by 4 categories of indicators, a total of almost 600 dimensional evaluation system.At present, the Libra open source large model evaluation system has been open to the public for registration applications. In terms of hardware, it supports various chip architectures such as Nvidia, Huawei’s Ascend, Cambrian, and Kunlun Core, as well as various deep learning frameworks such as PyTorch.



visual large modelthe Zhiyuan Conference announced the multi-modal large model Emu that complements everything in the multi-modal sequence, the strongest billion-level visual basic model EVA, the most powerful open source CLIP model EVA-CLIP, and pioneered the contextual image learning technology path The general visual model Painter, the universal segmentation model of the horizon that divides everything, and the first zero-shot video editing method vid2vid-zero and other achievements.