DL Developer for Neuro Quality Group

Neuro is the search of the future, combining large generative model technologies with classical web search.

Our team is responsible for answer quality in e-commerce scenarios (everything related to products in Yandex search results). Scenarios for selecting and purchasing products are a significant part of the Search business, generating a substantial share of revenue. And our projects are R&D at the intersection of generative AI and e-commerce.

We improve model quality on complex tasks, enhance the quality of information extraction, conduct numerous experiments in the process — and as a result, millions of people use our models. You will work with Yandex's most advanced models and technologies, creating new products for the Russian market.

What tasks await you

Improving the quality of generative models Currently, models are good at handling high-volume tasks, but quality may be insufficient for narrow e-commerce scenarios. On such segments, there is significant room for improvement and experimentation with LLMs, information extraction, data, and response format. You will participate in every stage of alignment: SFT (Supervised fine-tuning), Reward model, RLHF (Reinforcement learning from human feedback).

Improving information extraction The internet is full of diverse information, and sometimes it's not of the best quality, while at other times reliable sources are hard to find. Furthermore, information about products is not limited to web documents found via search queries. You will refine information extraction for building answers and experiment with new sources.

Model inference compression and cost reduction Neuro is used by a huge number of people, so we need lighter models in production. You will experiment with architectures and model reduction methods to meet latency and response generation speed constraints.

Product integration You will create a high-tech, convenient, and useful product for the user.

More about ML at Yandex — in the Yandex for ML channel

We expect you to

• Know Python
• Have good knowledge of classical DL and NLP
• Have solved NLP tasks using transformers
• Understand how modern LLMs work, have solved applied tasks with them or have relevant research experience