Kuaishou open-sources recommender model and benchmark (github.com)

Kuaishou has unveiled OpenOneRec, an open-source framework aiming to merge traditional recommendation systems with Large Language Models (LLMs). This initiative is significant as it addresses common challenges in generative recommendation, particularly the limitations posed by isolated data silos and inadequate reasoning abilities in existing models. The framework includes RecIF-Bench, a comprehensive Recommendation Instruction-Following Benchmark featuring 100 million interactions from 200,000 users across diverse domains like short videos, ads, and products. Additionally, the OneRec-Foundation Models, available in 1.7B and 8B parameter versions, leverage the Qwen3 architecture and are enhanced with a substantial industrial dataset to improve performance. OpenOneRec’s full-stack pipeline offers tools for data processing, co-pretraining, and post-training, fostering reproducibility and scalability in research. With a hierarchical benchmark that facilitates nuanced evaluation—from semantic alignment to reasoning—the models achieve state-of-the-art results across various tasks. The framework's design allows it to treat items as distinct modalities through Itemic Tokens, enhancing the LLM's processing of interaction history. Notably, its performance metrics show significant improvements over existing approaches, making it a valuable asset for the AI/ML community, particularly for those focused on enhancing recommendation systems.