Description
We are GigaChat Alignment. We make the model useful and reliable: SFT/DPO, distillation into small models, LoRA service, metrics and validation pipelines. We quickly test hypotheses, speed up training, and roll out improvements to production — first for internal clients, then for all of Russia.
Directions
Improving SFT / DPO: testing new training approaches, accelerating pipelines, generating new data, distilling knowledge from large LLMs into small ones.
Developing quality metrics for GigaChat, for example, by evaluating its ability to solve problems from international-level olympiads. Developing an internal LLM-AS-A-JUDGE.
Developing a LoRA training service for GigaChat and GigaEmbeder. Improving the stability and reproducibility of runs, creating validation and data generation pipelines using LLMs.
For these roles, we are looking for a talented NLP Engineer, with whom we will together improve and develop GigaChat. For all these experiments, we have a cluster with a large number of A/H 100s.
Responsibilities
- distributed training of models at the SFT/DPO stages, model distillation
- conducting research in the field of SFT/DPO to improve training quality and accelerate the process
- assisting in automating end-to-end model training processes and measuring their quality
- active interaction with the online-rl team to improve cold-start reasoning metrics
- analysis of training datasets, identifying relationships and the influence of data on final metrics.
Requirements
- higher education from a top university in Russia or abroad
- strong knowledge of algorithms and data structures
- experience in training LLMs (SFT, DPO)
- experience in configuring local inference (SGLang, vLLM, TRTLLM)
- understanding of how Python works under the hood
- ability to analyze scientific papers, reproduce them
- experience with distributed systems (Ray, Dask, OpenMPI)
- strong knowledge and experience with Linux, Bash
- strong knowledge of PyTorch
Conditions
- comfortable modern office - Kutuzovskaya metro station
- annual salary review, annual bonus
- corporate gym and relaxation areas
- more than 400 educational programs from SberUniversity for professional and career development
- extended voluntary health insurance, preferential insurance for family and corporate pension program
- flexible discount on mortgage loans, equal to 1/3 of the Central Bank's key rate
- free SberPrime+ subscription, discounts on products from partner companies
- referral bonus for recommending friends to the Sber team
- corporate pension program.
