MiniOneRec

Why we included this project

MiniOneRec is the first fully open-source framework to cover the whole generative recommendation pipeline, from turning products into discrete semantic tokens to training a small LLM to predict the next item. It uses a three-level RQ-VAE for that tokenization, then fine-tunes the model on next-item prediction, first with supervised fine-tuning and then a GRPO-based reinforcement learning stage aimed at ranking. The repo also includes offline Top-K evaluation scripts and runnable configs for Amazon datasets, plus constrained beam search that guarantees valid item outputs. If you've tried to reproduce papers like OneRec or GPR and hit gaps in the public code, this is a practical reference. Teams thinking about swapping a classic two-stage recommender for something language-model-based can treat it as a working starting point.