About the team: Our team is focused on training our own foundational LLMs and integrating them into Avito's diverse business scenarios. The team's goal is to provide all products with optimal models in terms of both response quality and cost-efficiency. We adapt the best open-source solutions for the Russian language and Avito's domain specifics using Continual Pre-training and tokenizer substitution. For continuous model improvement, we constantly research new methods, pipelines, including RL and synthetic data, and datasets. To stay at the forefront of technology, we hold regular LLM seminars where we discuss recent articles and AI trends.

Our technical focus today: — Building an optimal and scalable pipeline for LLM pre-training and fine-tuning. — Inference optimization: working with modern engines, quantization, speculative decoding, and other acceleration methods.

Examples of future tasks: — Scaling the A-vibe foundational model: researching new architectures and training approaches. — Optimizing the speed and cost of inference for high-load models in production. — Designing and assisting in the development of internal LLM platforms and multi-agent systems. — Building a robust infrastructure for model quality evaluation (evaluation).

We expect you to: — Have a solid understanding of machine learning and deep learning fundamentals, know the Transformer architecture and basic approaches to working with NLP models. — Have practical experience in training or fine-tuning neural network models and working with datasets, metrics, and validation. — Be proficient in Python, comfortable with PyTorch, and familiar with basic engineering practices: Git, testing, Code Review. — Be able to correctly set up experiments, compare approaches, and explain the reasons for specific results. — Be able to independently bring tasks with clear objectives to completion, proactively highlight risks, and seek help when needed. — Be interested in LLMs, read technical materials, and be ready to quickly master new tools and approaches.

It would be great if you: — Have at least 3 years of experience in implementing and operating ML solutions. — Have worked with Hugging Face Transformers, PEFT/LoRA, or other LLM fine-tuning tools. — Have experience with modern NLP models or have participated in their production deployment. — Are familiar with distributed training or inference optimization tools: DeepSpeed, vLLM, SGLang, TensorRT-LLM, and similar. — Have used MLOps tools for experiment tracking and data versioning: Weights & Biases, MLflow, DVC, and others. — Have participated in research projects, open-source development, or 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: voluntary medical insurance with dental care from day one, and on-site access to a therapist and masseuse. — Flexibility: the option to work remotely or from offices in 4 Russian cities.