Yandex introduced the third generation YandexGPT Lite
By ForGeeks Yandex has launched YandexGPT 3 Lite, a lightweight version of its own third-generation generative neural network. It will be available to Yandex Cloud customers via API. The new model is useful in scenarios where speed of response is important: for example, it can be used in chatbots, for spell checking or data analysis. It is the optimal commercial Yandex model for routine tasks in terms of price and quality.
YandexGPT 3 Lite is suitable for different types of businesses, from small companies to large organizations. It can be used to optimize tasks such as chat and phone consultations with customers, preparing answers for the helpdesk, creating marketing materials, or digesting work meetings. Large companies with complex business processes and high information flow can use it to analyze data for decision-making.
The new model outperforms YandexGPT 2 Lite, a lightweight model of the previous generation, in many ways. In the YaMMLU_ru test (the Russian-language version of the international MMLU benchmark test), the new model yields 6 p.p. more correct answers than the previous-generation model.
The new model gives 6 p.p. more correct answers than the previous-generation model.
The models were also compared using Side by Side methodology: neural networks answered the same questions and experts chose the best answer. On average, YandexGPT 3 Lite answered better than YandexGPT 2 Lite 68% of the time.
Experts also evaluated how well the new model handles categorization, content generation, question answering and other basic types of business tasks. Here’s what the test results look like:
The new model also makes fewer spelling and factual errors than the second-generation YandexGPT 2 Lite.
To create the new model, the developers have improved all stages of training. In particular, they improved the selection of data for the pretraining phase, increasing the proportion of useful information. In addition, they used curriculum learning technology to increase the complexity of the data in stages. In the second stage of training (alignment, or model alignment), which includes reinforcement learning, we improved the model for assessing the quality of neural network responses. In addition, Grouped Query Attention technology was added to the neural network architecture – it accelerates data processing without loss of quality.YandexGPT 3 Lite can be integrated into your products via API in the service Foundation Models. The new model will replace the previous one within a month, but you can try it out now. The cost of using YandexGPT 3 Lite is 20 pennies per thousand tokens. New Yandex Cloud users will be able to test it in demo mode for free.
.
