Description
The Research Center is engaged in creating State-of-the-Art artificial intelligence technologies for the analysis of financial data and their application for AI transformation in the Risk Block. Our task is to develop and train a multimodal reasoning model that combines textual data and customer digital footprints to improve the quality of solutions for banking tasks.
Responsibilities
- generate and clean synthetic data using LLM and LLM-as-a-Judge approaches
- train a large multimodal LLM (text + customer digital footprints)
- apply RL approaches to improve the model's reasoning abilities
- validate the developed solution on modern benchmarks in terms of accuracy and correctness of reasoning
- evaluate solutions on real-world data.
Hardware and Data:
- dedicated cluster of 500+ A100 80GB GPUs
- Spark/Hadoop for working with Big Data
- access to all banking data necessary for building the solution.
Requirements
- at least 3 years of work experience
- deep knowledge of ML/DL (in particular, transformers, large language models)
- confident experience in training LLM models
- solid knowledge of Python: torch, transformers, peft
- English level B2+.
Will be a plus:
- experience working with multimodal models
- knowledge of RL methods for LLM (RLHF, DPO, GRPO, RLVR, GSPO)
- experience in distributed training of large models (FSDP, ZeRO, Mixed-precision, Tensor/Pipeline Parallelism)
- publications in top-tier journals (Q1) or conferences (A-A*).
Conditions
- comfortable modern office at Kutuzovsky Prospect or Volgogradsky Prospect metro stations
- work format - office-based
- annual bonus
- corporate gym and recreation areas
- access to more than 400 educational programs from SberUniversity for professional and career development
- regular meetups and a well-developed DS-community
- extended health insurance, preferential insurance for family, and a corporate pension program
- flexible mortgage discount equal to 1/3 of the Central Bank key rate
- free SberPrime+ subscription, discounts on products from partner companies
- referral bonus for recommending friends to join the Sber team.
