Generates each subsequent word in a sentence so that the text is coherent and grammatically correct.
More
YaLM (Yet another Language Model), inspired by Open AI's GPT-3 and other language models on the Transformer architecture.
The task is to generate each subsequent word in a sentence to make the text coherent and grammatically correct. During training, the model evaluates each predicted word: for example, it decides whether the word run" or the word "frame" can go after "Mom washed...".To memorize all the rules of the language and pick up the right words Balabob is helped by the parameters embedded in the language model YaLM, which change depending on whether the word is predicted correctly or incorrectly.
You can compare them to little levers, each of which must be turned in different directions to start the mechanism.
The YaLM family of language models has between 1 and 13 billion such levers, while Balabob uses a model with 3 billion.
Terabytes of data are used to make the texts written by Balaboba not only grammatically correct but also lexically diverse.
YaLM is trained on a fraction of pages indexed by Yandex on Runet, including not only Wikipedia, news articles and books, but also open entries by users of social networks and forums.
Not to overload the model, from a sample cleaned repetitive, incomplete and unnatural texts.
But the main feature of YaLM is the ability to learn new things from just a few examples: for a language model to write meaningful movie reviews, dinner party toasts or conspiracy theories, five to several dozen examples of how these texts should be written are enough.
This is exactly what you can observe in the choice of stylization: for example, to teach Balabob to generate simple advertising slogans for any object, he was "fed" several famous examples, among which, of course, Finds Everything.