Yandex has released an updated version of its language model - YandexGPT 2

At the conference Practical ML Conf Yandex introduced the second generation of the YandexGPT language model. In this version, the developers have increased the number of model parameters, supplemented the initial data set with new query examples, and tightened control over the quality of information on which the model is trained.

The new version is able to adapt answers to different audiences and rewrite them in other styles and genres. Thanks to this, a neural network with YandexGPT 2 will be able to explain a complex term to a child in simple language or retell the director’s decree in the form of an understandable letter for distribution to employees.

The model is also able to work with texts up to a thousand characters long, sent by the user: analyze them, structure information and answer questions on “read”.

Also YandexGPT 2 was compared with the first version, forcing them to respond to the same set of queries. After that, the results were checked for accuracy, completeness, factual correctness, grammatical and stylistic correctness. It turned out that the second version copes much better with a number of requests. In particular, text generation, retelling and analysis of text, idea generation, styling for the audience and characters, and answers to factual queries have improved.

YandexGPT 2 is already available in Alice's "Let's Think" mode in the Yandex app, Yandex Stations, TVs with Alice, Yandex Browser, on the search results page, and on ya.ru.