Tatiana Shavrina, PhD

Llama team

Leading LLM development for MyAI

Working at AIRI as a director of key research projects, responsible for two core research teams: NLP and Multimodal models. Managing 2 team leads and Proposing research for year-long initiatives in large language models, multilingual models, model interpretation and evaluation.

Working as an AI Research Leader for Sber within the Experimental Machine Learning Department responsible for 6 teams: AI Prototypes, Applied Research, Large Transformer Models and Architectures in Neural Networks Manage 5+ Agile R&D teams with 30+ employees, value stream ownership in NLP and multimodal ML. Built and designed organizational structure for AI Prototyping and Research following SCRUM methodology, managing sprint cycles and stakeholder engagement Contributed 20+ research initiatives in AI focusing on natural language processing, some examples include: Interpretable NLP (INLP), Efficient Benchmarking in NLP, NLP applications to field linguistics, and Artificial Text Detection task Contributed the ruGPT-3 model to the Digital Assistant: implemented its dialog fine-tune to production, leading to a 23% increase in retention over previous retrieval solution. Created the release train to implement neural-based skills to the Digital Assistant. One of the skills, text-to-image generation model (ruDALL-E) has brought 40k+ new registered users in the first week and 2.2 million unique users over all time, 125 million user requests. Set up the B2B model zoo for AI cloud clients, bringing 30+ corporate clients. Lead language model evaluation leaderboard, 1500+ new models have been submitted by the community leading to a more than 60% increase in open-source models’ performance over time.

- NLP-based prototyping: in-platform spell-check, text autocompletion, NER-platform, few-shot text classification platform (B2C). - Using static word vector models like word2vec, fasttext, as well as emerging ULMfit, Bert models for few-shot text classification tasks, such as the assignment of payments, and cataloguing of product cards. - Implemented NER-prototype in the corporate email system, automatically fills in the dates, times and participants of the meetings, as well as the subject of the meeting. - Pet-project: created “Taiga” - 6 billion words open-source web-corpus for machine learning

Project coordinator: managing work of the developers, students in practice, testers, checking working timeline according to the deadlines, content-managing the pages of the project in social networks, updating the project website.

Conducted a full pipeline of adding new texts and resources to the corpus, including source validation, OCR, error correction, automatic text annotation, deduplication, html formatting.

Methodology of science in NLP

Department of theoretical and applied linguistics

Summer courses of Finnish language and culture