The name of the base system (platform): | YaLM 100B (neural network) |
Developers: | Yandex |
Date of the premiere of the system: | August 2022 |
Branches: | Internet services |
2022: Service Launch
At the end of August 2022, Yandex launched a bilingual version of the Balaboba text generator. The user can write one or two words in Russian or English and choose one of the styles - and "Balaboba" will create a meaningful text on any topic, similar to the texts from the Internet on which the model studied.
To make the text coherent and grammatically correct, the model generates it sequentially from the words and estimates how correctly the predicted word is. For example, after "Mom Soap..." the word "run" or the word "frame" can go.
"Balaboba" can write a short story, come up with a recipe, instruction or folk wisdom. And if you enter the name of the film, "Balaboba" will write a plot for it. Texts can be used at your discretion. For example, describe products in an online store, create text for advertising, etc.
Balaboba generates texts using Yandex's YaLM language model, which solves problems related to natural language processing. For example, YaLM models help Alice maintain a conversation, define question topics in Kew, improve order descriptions for Services, and generate cards for quick answers in a search engine. Also, YaLM language models look for key video points, generate advertisements and site descriptions (snippets).
Memorizing the rules of the language and selecting the appropriate words "Balabobe" allow the parameters embedded within the model, which change depending on whether the word is correctly or incorrectly predicted. The YaLM family of language models can have from 1 billion to 100 billion parameters.
The largest bilingual model YaLM 100B for 100 billion Yandex parameters was publicly available earlier in 2022. Balaboba uses its lightweight version of 3 billion parameters. The model was trained on terabytes of texts of the English and Russian-speaking Internet in equal proportion.