ZooWork-ShopRanker: An Open, Preference-Aligned E-Commerce Reranker

This work introduces ZooWork-ShopRanker, a family of e-commerce rerankers in 0.6B, 4B, and 8B sizes designed to handle ranking decisions driven by product constraints and user preferences. Since real-world search traffic lacks clean pairwise labels, the authors utilize a panel of reasoning large language models as a preference oracle with position-debiased judgments. The aligned 8B flagship model then acts as a distillation teacher for the 4B and 0.6B variants. To measure progress, they release ShopRank-Bench, a contamination-limited benchmark containing approximately 10,000 private-traffic preference pairs.

The ZooWork-ShopRanker family comprises 0.6B, 4B, and 8B models aligned using judge-labeled shopping preferences.

The aligned 8B flagship model serves as a distillation teacher for the efficient 4B and 0.6B models.

ShopRank-Bench is introduced as a contamination-limited benchmark consisting of ~10,000 private-traffic preference pairs.