Method

QLoRA

LoRA on a 4-bit quantized base model.

7 cases from 7 companies apply QLoRA. 1 was checked in an online A/B test; 2 report results in production.

Typical tasks: Rank search results, Reduce ML cost, Label data with LLMs, Build an assistant over internal docs and 1 more. Companies: Sber, VTB, Pinterest, Mercari, CFT and 2 more. Often combined with: RAG, Supervised fine-tuning, Knowledge distillation, BM25, Reinforcement learning and 16 more.