Senior LLM Developer for the Neuro Team

Neuro is a major technology project that combines LLMs, LLM agents, assistants, and web document search. Our goal with it is to build the search of the future — smarter, deeper, more useful, and more proactive. One of the most crucial properties of such a system is answer reliability, which implies factual accuracy and verification based on data from reliable sources. We are developing approaches to ensure these properties in our generative models and are looking for an LLM developer to strengthen this area.

What tasks await you

Creating and improving reward models For training generative models using RL, it is critically important to have good rewards that can automatically assess the conformity of answers to required properties. We train discriminative and generative reward models, using all available modern methods to improve their quality: SFT, RL, reasoning. This work directly impacts the improvement of answer quality for users.

Training LLM assessors We aim to make the process of evaluating answer quality more efficient and accurate. You will work on creating LLM agents that will assist humans with fact-checking and data labeling, thereby making these processes faster, cheaper, and even higher quality.

Compression and cost reduction of LLM assessors and reward models Training strong LLMs requires large amounts of data, which means we need to run many labeling and training processes. You will experiment with architectures and methods for model reduction and inference acceleration, so that all our pipelines converge faster.

Learn more about Alice AI

We expect that you

• Know Python
• Have a good understanding 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