Senior LLM Developer at Neuro

We are developing Neuro — a system of models for constructing LLM responses using search as an information source and for answer enrichment. We have implemented such a system in Yandex Search and now face a new challenge: learning to solve complex search-based scenarios that arise in conversations with Alice.

Imagine the query "create a travel plan to see orcas in the summer". To build a good response to it, multi-stage search, the use of images and data from hotels, a good structure, and high reliability are needed. And there are hundreds of thousands of such queries across different aspects.

Our team is responsible for Neuro's Align models: we train SFT, research the newest RL algorithms, maximize rewards, implement long contexts and MoE. The part with model optimizations is also important: Neuro is used by tens of millions of people, and to handle such a load, we constantly need to find and implement new optimization methods. We are currently facing new challenges and are looking for a strong DL developer for our team, who knows LLMs and knows how to write code.

What tasks await you

Research of RL approaches Since Neuro should not only write texts but also use images, videos, appropriate structure, we use dozens of reward models to optimize different aspects. Currently, our main methods are DPO and CE-RL, but they are not perfect, and we want to learn more efficiently. You will need to read papers, implement new methods, and also improve the current ones.

Model optimizations Rolling out Neuro to millions of users is very expensive, and when it comes to multi-stage searches, the numbers become truly astronomical. You will need to research, implement, and sometimes debug quantizations, distillations, spec-decks. And then deploy them without quality loss.

Search for new directions When you make a new product, many challenges arise, and it's important to find growth points among them, which later turn into tasks. For example, we lack a longer context, we lack the ability to "see" images. When such a growth point is found, you need to turn it into a task, agree with related teams on joint actions, and bring this task to completion not only in development matters but also in management. Such tasks perfectly help grow towards becoming a lead or tech lead, and they arise regularly.

More about Alice AI

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

We expect you to

• Understand how LLMs are structured, how they are trained, and what degrees of freedom they have
• Know classical DL and ML
• Write code in Python
• Read papers, are ready for complex challenges and ambitious tasks

Will be a plus

• Have worked with RL
• Have implemented optimization methods or model production inference