Senior Data Scientist for the LLM Team

About the Team

Our team is engaged in training our own foundational LLM and applying it to various business tasks at Avito.

To develop the foundational model, we adapt the best open-source models for the Russian language and the Avito domain using Continual Pre-training and tokenizer replacement. You can read about this in articles on Habr:

→ How we taught Mistral 7B Russian and adapted it for listings

→ How we at Avito made our own LLM — A-vibe

To improve the model, we research new methods and datasets. And to ensure everyone on the team is on the same page, we hold LLM seminars where we discuss the most interesting articles.

Already, with the help of LLMs, we have solved many interesting and useful tasks for Avito. Here are a few examples of products where we have already managed to implement LLMs:

Description Generation. In some Avito categories, it is no longer necessary to prepare a listing description yourself—you can take a text generated by the LLM.

Modification of Autoteka Reports. Avito receives data for them from partners who often use wording and abbreviations that are unclear to ordinary people. We trained an LLM to decipher them.

Summarization of Support Agent Chats. When an agent cannot solve a problem, they can transfer it to a more experienced colleague. To do this, they need to briefly describe the content of the chat with the user. Now the LLM can do this.

Modification of Support Agent Messages. We trained an LLM to rephrase some support agent messages to make them more empathetic and correct mistakes.

Suggestions in the Messenger. When writing a message on Avito, you may encounter pop-up suggestions from the LLM—they help you communicate more conveniently and quickly in the chat.

Examples of Future Tasks:

• research articles and improve the foundational model;

• optimize the inference speed of models;

• assist in the development of platform LLM solutions.

We Expect You to:

• understand how the main ML algorithms work (from decision trees to transformers);

• have experience working with and deploying ML models in production;

• know Python;

• understand how LLMs work, follow AI trends;

• have worked with modern NLP models.

It would be great if you:

• have achieved high places in machine learning competitions;

• have used experiment tracking tools: Weights & Biases, MLflow, DVC, etc.

Working With Us Means:

• the opportunity to improve the experience of millions of users;

• interesting and challenging tasks on a large scale;

• a strong team that is always ready to help;

• the opportunity to study and try new things, with powerful hardware for this;

• a training budget that can be spent on courses or professional literature;

• care for your health: from day one you will have health insurance including dentistry, a therapist and a masseuse are available at the office;

• the opportunity to work remotely or from offices in four cities in Russia.