Problem
Sellers on a classifieds marketplace often leave the description empty or useless: a single dot, an emoji, "new dress, excellent condition". Clothing buyers need measurements (waist and inseam for jeans; skirt length, bust, waist and sleeve for dresses) that the card attributes do not carry. Without them the buyer must message the seller before deciding, which adds an extra step. Avito's LLM team wanted descriptions that are informative and that sellers actually accept, in the Fashion category.
Example from the article: a real listing whose description is just "." and another with only a "like" emoji.
Solution in brief
A small image-to-text encoder-decoder reads the main photo, the attributes and the title, and writes the description. It is trained on Avito listings filtered by business-defined criteria of a good description. Every number in the output becomes a placeholder, so the model cannot invent measurements, and pre- and post-filters guard the output. In an A/B test on private sellers, orders with Avito Delivery rose 1.7%.
Questions this case answers
- How to generate a short descriptive text from an image and structured attributes without the model inventing numeric facts that the input cannot show
- How to validate a generated-text feature with domain reviewers before running an online test
- How to auto-generate listing descriptions from a photo and structured attributes
- How to prevent a generative model from hallucinating measurements in product descriptions
Tools and methods
- BLIP-like encoder-decoder — image to encoder, text prompt to decoder; open model trained on COCO captions translated into Russian
- Business-defined training sample — good and bad description features collected with the Fashion team
- Pre- and post-filters — skip unanswerable requests; reject rule-breaking generations
- Number placeholders — blanks for the seller to fill
- Hugging Face Transformers, Kubernetes, Grafana — serving and monitoring
Google or a one-time email code. No password.