Avito has launched the 'Personalized Ranking | Avito ML cup 2026' competition on the Open Data Science platform: for each recommendation feed, you must predict a numerical score for each candidate listing so that those with which the user will perform a target action rank higher. Solutions are evaluated using the NDCG metric, the prize pool is 1.5 million ₽, the data and two starting baselines have already been published, and you can register with a team until December 8, 2026.
What happened
Avito opened the competition on the ODS platform in the fall of 2026. The data is divided into train/val/test, each feed_id contains at least 5 candidates, and the final score is calculated as NDCG averaged across feeds. Participants have access to 5 numerical features, categorical features, 128-dimensional user and listing embeddings, and a history of up to 256 user events. Solution submissions open on October 28, 2026 — 5 submissions per day per team. Registration and team formation (up to 4 people) will last until December 8, and the final is scheduled for December 9, 2026. The prize pool is distributed by rank: 750,000 ₽ for first place, 450,000 ₽ for second, and 300,000 ₽ for third. Two baselines are provided to get started — CatBoostRanker (YetiRank) and Transformer.
Context
Personalized ranking is a key task for marketplace recommendation systems: the order of listings in the feed determines what the user will see and interact with. Companies almost never publish industrial data for such feeds, so open datasets with vector representations and long behavioral histories are rare in the industry. The competitive format is well-established: for a fixed prize pool, the organizer receives numerous third-party solutions on their real data, while the community gets an open leaderboard and a common point of comparison for approaches. Evaluation on a hidden test, rather than a public leaderboard, is considered best practice for competitions: it reduces the risk of overfitting solutions to the visible set. The two baselines set a difficulty corridor: gradient boosting serves as a strong tabular benchmark, while the Transformer invites modeling event history as a sequence.
Why this matters for the industry
For the industry, this is a rare case where a major marketplace opens anonymized data from a production recommendation feed and sets a public benchmark for working with embeddings and personalized ranking. The dataset with vector representations and a history of up to 256 events is material for sequence models and methods that aggregate behavioral history, not just for gradient boosting on tabular features. For Avito itself, this is a cheap R&D and hiring mechanism: for 1.5 million ₽, the company gets the work of hundreds of teams on its data, and the market gets a public leaderboard as a ready-made funnel for recsys talent. After the final, as is usually the case with major competitions, a public layer of solutions and analyses will accumulate: combinations of tabular features with embeddings, schemes for working with long event histories, and validation practices for feed-based metrics will become a comparable industry reference point.
Why this matters for users
For ML specialists, this is free access to an industrial dataset and ready-made starting models: in the evening, you can spin up a working CatBoostRanker (YetiRank) or Transformer baseline on the provided data. The first reasonable task is to run both baselines and build local validation that reproduces the NDCG averaging across feeds: with a limit of 5 submissions per day, all iterative work should be done on your own held-out samples, not on submissions. You can participate in teams of up to 4 people, registration is open until December 8, 2026, so you still have a month to enter the fight for the prize pool. This is a training ground, not a ready-made tool: the solution does not bring a direct effect to working products, but it provides an industrial task for practicing skills, and results in the open leaderboard are visible to employers.
What is currently unknown / limitations
The strength of the baselines is unknown: CatBoostRanker (YetiRank) and Transformer have not been publicly compared, and it cannot be asserted in advance that modeling a history of 256 events will yield an improvement over gradient boosting — this is a hypothesis to be tested on the leaderboard. The final evaluation is conducted on a hidden test, so local validation may differ from the official result. The specific target action and details of data anonymization are not disclosed in the published materials; it is also unknown whether analyses of winning solutions will be published.
Sources
- Personalized Ranking | Avito ML cup 2026 — Open Data Science
- FAQ — Personalized Ranking | Avito ML cup 2026
- Fall Avito ML Cup: ranking of recommendation feeds — Hacklist
Author
Look at AI, editorial team