Author: Albert Akopyan, Core LLM team, Avito.tech.
A talk about how Avito post-trains its own base model built on open weights: a custom tokenizer, SFT and RL with verifiable rewards (RLVR) in three domains, a change of training framework when moving to MoE, and engineering techniques for GRPO.
Continue reading — it's free
Google or a one-time email code. No password.