Data Scientist for the LLM Core team

About the team:** Our team trains its own foundational LLMs and integrates them into Avito's business scenarios. The team's goal is to provide products with models that meet quality and cost-of-application requirements. We adapt open-source solutions for the Russian language and Avito's domain specifics using continual pre-training and tokenizer replacement. To develop models, the team researches new methods and pipelines, including RL and synthetic data, as well as datasets. We conduct regular LLM seminars where we analyze articles and AI trends. The team's technical focus is building a scalable pipeline for LLM pre-training and fine-tuning, as well as inference optimization: working with modern engines, quantization, speculative decoding, and other acceleration methods. We are looking for a Data Scientist to join the LLM Core team. You will be responsible for creating and developing foundational language models for Avito's audience: from formulating research hypotheses and conducting experiments to deploying solutions into production. In this role, you will influence the model architecture, training pipelines, and the quality of the company's LLM stack.

Examples of future tasks: — Formulate and test hypotheses in the areas of pre-training, continual pre-training, post-training, RL, synthetic data, and model architectures; — Independently lead complex research and engineering tasks: from setting goals, choosing metrics, and designing experiments to implementation, result analysis, and deployment; — Scale training and inference: profile bottlenecks, improve the efficiency of distributed training and inference, work with quantization, speculative decoding, and modern inference engines; — Build reliable quality assessment procedures, conduct error analysis, identify regressions, and design datasets and metrics for comparing models by quality, latency, and cost; — Write production code in Python and PyTorch, conduct code reviews, improve experiment reproducibility, reliability, and maintainability of ML solutions; — Share knowledge, mentor less experienced colleagues, participate in architectural discussions and internal LLM seminars. The role does not involve mandatory team management; — Scale the foundational A-vibe model: research new architectures and training approaches; — Optimize the speed and cost of inference for high-load models in production; — Design and help develop internal LLM platforms and multi-agent systems; — Build a reliable infrastructure for assessing model quality.

We expect you to have: — At least 3 years of experience in implementing and operating ML solutions; — Deep understanding of modern ML algorithms, Transformers, and LLM training principles; — Practical experience in training, adapting, or deploying NLP models and the ability to turn ideas into testable experiments; — The ability to work independently with uncertainty: decompose problems, choose metrics, analyze results, and make technical decisions; — Proficiency in Python, strong knowledge of PyTorch, and software engineering practices; — Understanding of the trade-offs between quality, latency, throughput, and cost, and consideration of these factors when choosing solutions.

It would be great if you have: — Experience with continual pre-training, tokenizer replacement or extension, SFT, LoRA, or RL approaches to post-training; — Experience with distributed training or inference optimization and familiarity with tools like DeepSpeed, vLLM, SGLang, or TensorRT-LLM; — Experience building quality assessment pipelines, synthetic datasets, or automated model error analysis systems; — Experience with MLOps tools for tracking experiments and versioning data and models: Weights & Biases, MLflow, DVC, or similar; — Published research, participation in open-source projects, or high rankings in machine learning competitions.

Working with us means: — The opportunity to influence the product: your decisions will improve the experience of millions of users; — Complex technological challenges at a large scale; — Powerful hardware: access to high-performance GPU clusters for experiments; — A strong community: a team of experts with whom you can discuss ideas and approaches; — Development: a budget for training, courses, conferences, and professional literature; — Health and well-being: comprehensive medical insurance with dental coverage from day one, on-site access to a general practitioner and massage therapist; — Flexibility: the option to work remotely or from offices in 4 Russian cities.