Sber has opened access to the SIRIN library. It identifies fictional and unconfirmed facts in the responses of language models. It is based on Sber's research in the field of neural network "hallucination" detection.

SIRIN analyzes the model's response as a whole or individual text fragments and highlights areas where errors or unconfirmed information may be present. The tool pre-evaluates whether the AI agent has enough information to answer. If there is insufficient data, the model can request additional information or refuse to provide an unconfirmed answer altogether.

Scientists from the company's team have developed metamodels that increase the accuracy of error detection by almost 30% — all based on only 250 training examples.

The solution combines various methods for verifying language model responses: developers can compare them, combine them, and choose the appropriate option for their task. The approach has proven itself within Sber in customer service applications, and now the library has been opened to all companies.

"Hallucinations" in AI refer to situations where a language model confidently presents fictional information as fact. The neural network does not "recall" data but predicts the most probable continuation of the text — and can with equal confidence provide a correct answer or invent non-existent events, quotes, or figures. Such responses appear plausible, making it difficult to recognize an error without verification.