RocketHuntSmart job search
Яндекс
Russia5 days ago
Reach out directly about this role
Employment
Full-time
By company and country
AI writes it from your resume. You only have to send it.
We are looking for an ML developer to improve the quality of Alice AI LLM pre-training. This area is responsible for the intelligence of the base model and data quality.
The LLM race is accelerating: models are solving problems in mathematics, code, agents, and medicine that recently required human intervention. At the same time, major improvements in LLMs come from updating and scaling pre-training: it is at this stage that the fundamental abilities of the model are laid, on which quality depends in all subsequent contexts.
We believe that the winner in this race is the one who improves pre-trains more effectively and faster, and we are investing in this direction: we research ways to improve model quality, conduct computationally intensive experiments, and process and synthesize vast amounts of data.
The result of our work is the Alice AI family of models, available in the service of the same name. By making our model smarter, you will improve the main and most widely used AI assistant in Russia, which is already used by over 24 million people weekly.
What we have already done: Our current pre-train models have already shown successful results in previous Alice releases. For example, we wrote on Habr that we demonstrate the best quality on factual and cultural slices, measurements, and more. We discussed the research that formed the basis of these improvements at Practical ML Conf 2025.
What will you be doing? Model reasoning capabilities have a proven benefit in areas such as mathematics, code, and agents, but leading players are achieving good quality gains through reasoning on a broader range of tasks, from finance to computer games. Our goal is to make Alice's reasoning model the strongest in the Russian Federation and beyond.
Experiments with Pre-train Models Conduct and analyze computationally intensive experiments with the pre-train stage. Test hypotheses about the sources of reasoning ability gains within various aspects of intelligence. Designing and selecting the experimental setup is a creative stage.
Data Pipelines Build efficient pipelines for collecting, synthesizing, processing, and filtering training data in new domains. The constructed pipelines will need to be scaled and made more efficient.
Modifying the Training Process Modify the model training process to improve reasoning tasks. Implement changes in the training pipeline and evaluate their impact on quality.
Expert Models Train verifier models to check the quality of data and responses. Train models to generate training data that are SOTA in their narrow domain. Train filter models to discard low-quality data.
Scaling Research Investigate the scaling laws of model quality in terms of data and computation. Determine how scaling affects the growth of reasoning abilities in different domains.
More about ML at Yandex — in the Yandex for ML channel
