Yandex Introduces Lighter Version of Its Third-Generation AI Model: YandexGPT 3 Lite

Yandex, the multinational technology company, has officially unveiled a lighter version of its third-generation generative AI model, YandexGPT 3 Lite. This AI service is available to Yandex Cloud customers through its respective API and is versatile for several use cases, such as in chatbots or for spell-checking and data analysis.

Applicability in Business

The company highlighted that the new AI model version has a wide range of business applicabilities. The use of YandexGPT 3 Lite can streamline the execution of diverse tasks such as customer consultation via phone or chat, response preparation for customer support, and marketing material generation. More complex organizations managing vast data flows and business processes may find the algorithm useful for information analysis.

Performance and Testing

According to Yandex developers, YandexGPT 3 Lite excels in many parameters over its predecessor AI model. During the language model testing in YaMMLU_ru (the Russian version of the international benchmark test MMLU), it was established that YandexGPT 3 Lite delivers 6% more accurate answers than the YandexGPT 2 Lite model.

Furthermore, when the models were compared using the Side by Side method, in which they answer identical questions while the best response is chosen by an expert group, YandexGPT 3 Lite outperformed in 68% of cases. The testing process also evaluated how well YandexGPT 3 Lite handles classification tasks, content generation, question-answering, among others. Moreover, the new algorithm causes fewer spelling and factual errors compared to YandexGPT 2 Lite.

Creation Process and Improvements

Throughout the creation process of the new AI model, developers enhanced every training stage. The data selection for the initial training phase was improved, leading to an increased share of useful information. Additionally, the Curriculum Learning technology was employed for the gradual complication of data. The second training phase, which involves reinforcement learning, saw a bettered model for assessing the quality of the algorithm’s responses. Furthermore, the AI network architecture incorporated a Grouped Query Attention technology to speed up data processing without quality loss.

Costs and Implementation

The use of YandexGPT 3 Lite is priced at 20 cents for every 1,000 tokens. New Yandex Cloud users can trial the AI service for free in demo mode. This new model is set to replace the previous algorithm version within a month.