Problem
Wildberries moderates marketplace product cards (images and text) against rules such as weapons, 18+, alcohol and drugs, and the rules differ by country. About 15,000 cards per minute are updated and need near-instant decisions. Full re-moderation runs (rescores) add bursts of millions of cards. A single VLM that reads everything is too slow and non-deterministic, while plain keyword triggers and image classifiers produce false positives. Sellers also put products into the wrong category to dodge moderation.
Example from the article: a seller lists phenibut (fenibut) under "books" to bypass moderation, because trigger models and classifiers are tied to categories.
Solution in brief
Run fast deterministic checks first (Aho-Corasick triggers, regular expressions, ViT image classifiers) and call an LLM or VLM only on the flagged cards. The LLM/VLM either unbans false positives or, in silent mode, makes the ban decision itself. An LLM also checks the category, and embedding similarity routes new items to the right classifier.
Questions this case answers
- How to moderate a very high volume of items against rules that differ by region when a slow generative model cannot run on every item and deterministic filters must screen first
- How to find new items that resemble prohibited ones and route them to the right classifier when adversaries put items into the wrong category
- How to moderate marketplace product cards in near real time when an LLM or VLM alone is too slow
- How to cut false positives of keyword triggers and image classifiers with an LLM/VLM second check
Tools and methods
- Aho-Corasick trigger models — blacklists and whitelists per category, stored in S3
- Qwen 2.5-72B-GPTQ-Int4 and Qwen 2.5-VL-32B-AWQ on vLLM — LLM and VLM checks of flagged cards
- ViT classifiers, DALI and Triton — one fully fine-tuned model per image rule
- Embeddings and a vector DB — similarity of new cards to labeled prohibited items
Google or a one-time email code. No password.
