How to Speed Up Multimodal Labeling Fourfold Without Losing Quality: Training a Compact VLM on Complex Data and Domain Projects

Alexander Mandrov23 May 2026

Problem

Yandex Crowd labels data for 80+ Yandex products and launches 1300+ new projects a year. Some projects need 10 million+ tasks a month, and finding assessors for them gets harder. The team replaces assessors with a vision-language model (VLM), but the large model is too heavy: a stream of 18 million tasks a month needs 8 large and 55 small cards. The compact model needs 2 large and 11 small, four times fewer, but its average accuracy across projects is 64% against 70% for the large one.

Example from the article: determining the vehicle type from a photo, or deciding whether two photos show the same product.

Solution in brief

Fine-tune the compact model on the large model's reasoning, keeping only examples where that reasoning reaches the correct answer. Then keep only the hard ones by sampling the large model eight times per task (Polaris). Train separately per domain, a group of similar projects.

Questions this case answers
  • How to make a small vision-language model match a large one on a labeling task by training only on the large model's reasoning for the hardest examples
  • How to automate many heterogeneous labeling projects when the large model is too slow and needs too many accelerators
  • How to make a small VLM match a larger one on crowd labeling tasks
  • How to choose the most informative training examples by estimating their difficulty

Tools and methods

  • Alice AI VLM (32b) and Light (8b) — teacher and compact student
  • Reasoning distillation — reasoning generated by the teacher, filtered by correct answers
  • Polaris difficulty filtering — N sampled answers per task, hard tasks kept
  • Domain-specific supervised fine-tuning — one model per project group
Continue reading — it's free

Google or a one-time email code. No password.