GrocLM: Grocery Category Recommendation in E-Commerce with Large Language Models

Short summary

GROCLM is a fine-tuned LLM for grocery category recommendation that uses two-stage LoRA training to encode cyclical purchasing patterns and trie-based constrained decoding for valid outputs. In a live production restocking task, it achieved a 7.5% relative improvement in cart-adds per impression over strong baselines. The approach demonstrates practical integration of LLMs into structured recommendation systems at scale.

• •Fine-tuned LLM for grocery category recommendation using two-stage LoRA training
• •Trie-based constrained decoding ensures valid outputs over a predefined category space
• •7.5% relative improvement in cart-adds per impression in live production deployment

Generated with AI, which can make mistakes.