Data Engineer for VLM Training Data (GigaChat Vision)

Join Sber as a Data Engineer to work on VLM training data, focusing on data pipelines and ML team needs. Sber is a leading financial institution in Russia, providing a wide range of banking and financial services.

• •Gather and structure the ML team's data needs for training, fine-tuning, evaluation, and improvement of VLM.
• •Propose and implement ideas for data cleaning, filtering, deduplication, categorization, and generation pipelines.
• •Navigate modern practices for building datasets for Vision-Language Models: image-text pairs, synthetic data, filtering, quality scoring, data mixture design, dataset versioning.
• •Be responsible for the infrastructure for data storage and preparation, including:
• •importing data from various sources: production, Common Crawl, open-source datasets, generated data;
• •validating and controlling data quality;
• •storing and versioning datasets;
• •exporting data in formats suitable for model training.
• •Design and implement data processing pipelines at scale, including tens of billions of images.
• •Develop pipelines for generating synthetic data for training and improving VLM.
• •Collect statistics on data, build reports and visualizations for analyzing the composition, quality, and coverage of datasets.
• •Ensure reproducibility, observability, and reliability of data processes.
• •Work closely with ML engineers, researchers, and the infrastructure team.