Hugging Face Daily PapersSiqiao Xue, Shuxuan Liu, Ning Hu1 min readpaperadvanced
ZooWork-ShopRanker: An Open, Preference-Aligned E-Commerce Reranker
Summary
ZooWork-ShopRanker is a family of open e-commerce rerankers (0.6B, 4B, 8B) trained using preference labels generated by an ensemble of LLMs. The 8B model, aligned to judge-labeled shopping preferences, significantly outperforms strong open reranker baselines on a new benchmark, ShopRank-Bench.
- General web rerankers transfer imperfectly to e-commerce due to unique user preference signals.
- ZooWork-ShopRanker uses an ensemble of reasoning LLMs as a preference oracle to generate training labels.
- The 8B flagship model is distilled into more efficient 4B and 0.6B models for practical deployment.
- ShopRank-Bench, a new 10,000-pair benchmark, measures e-commerce preference alignment.
E-commerce search engineers can leverage these open models and the new benchmark to improve product ranking by incorporating nuanced user preferences beyond topical relevance.
8/10
