RocketHuntSmart job search
Reach out directly about this role
ML Engineer for the Recommendation Technologies Propagation Team
Our team researches and develops ML personalization models for recommendation services. We build transformers on top of user history; they are an important part of advertising technologies and recommendations in the Market.
Lately, the field of recommendation systems is becoming ever closer to NLP: during training, we separate the stages of pre-training and SFT, observe similar model scaling laws, and train on hundreds of GPUs. However, there are also significant differences: in services, the set of recommended entities changes dynamically, and the size of this set can reach on the order of 10^9. Besides, each user event carries much more information than a single text token.
Our goal is to combine the best of both worlds — RecSys and NLP — and improve specific products with our technologies.
Our R&D team develops cutting-edge recommendation technologies that are used across all of Yandex. We are looking for a strong ML engineer who will research new approaches in recommendations and bring them to a production-ready state. If you have a solid understanding of DL, are familiar with modern RecSys or NLP, and have experience deploying neural networks into production — we look forward to meeting you!
Pre-training or learning to replicate the logging policy In any mature service, a fairly high-quality recommendation system is already in place, so at the first stage the model must learn to reliably replicate existing recommendations. To achieve this, we experiment with data, architecture, losses, and other aspects.
SFT After pre-training, the model is trained on user feedback to rank relevant candidates and select the best among them. Among the open questions in this area are: what is the quality limit of the model for a specific task formulation, what do the scaling laws look like in different domains, and what further improvement paths are worth exploring.
Adapting models for production A major challenge for us is making models work in runtime under high loads of tens of thousands of RPS. We actively research architectural optimizations and use specialized inference frameworks, and sometimes even write our own CUDA kernels in Triton.
Opportunity for broad development As an R&D team, we are not limited to a single product or a single technology. If desired, you can delve into different services or try other approaches in recommendations.
More about ML at Yandex — in the channel Yandex for ML
3 years
Experience
Full-time
Employment
Middle
Grade
Data Science & ML
Specialization
IT & Tech
Industry
Corporation
Company Type
By job title
IT & Tech
Industry
Corporation
Company Type