# MiniOneRec

The first fully open-source generative-recommendation framework, providing an end-to-end pipeline of semantic ID construction, supervised fine-tuning and reinforcement learning, reproducing OneRec.

MiniOneRec is an open-source framework for scaling generative recommendation, reproducing the OneRec approach end to end. Its pipeline spans SID (semantic ID) construction, supervised fine-tuning, and recommendation-oriented reinforcement learning.

It includes multiple SID construction methods, such as RQ-Kmeans+ and constrained-RQ-Kmeans, a TS-Rec codebase implementing fine-grained semantics integration for LLM-based recommendation, scripts for processing datasets such as Amazon23, and a multi-GPU parallel text-to-embedding pipeline built on Accelerate. It supports freezing LLM parameters during fine-tuning to train only newly added embeddings.

It is released with a technical report and is also available through Hugging Face and Modelscope.

Website: https://github.com/AkaliKong/MiniOneRec