Method

RLHF

Aligning models with human preference feedback via PPO-style reinforcement learning.

6 cases from 2 companies apply RLHF. 2 were checked in an online A/B test.

Typical tasks: Label data with LLMs, Evaluate models, Build an assistant over internal docs. Companies: Rabota.ru, Wildberries. Often combined with: Supervised fine-tuning, Reward model, PPO, LoRA fine-tuning, Abstractive summarization and 20 more.