AkaliKong/MiniOneRec

Implements a complete generative recommendation pipeline using semantic item IDs (SIDs) derived from RQ-VAE quantization of text embeddings, paired with LLM-based next-token prediction. The framework combines supervised fine-tuning with language-alignment objectives, followed by GRPO-based reinforcement learning using constrained beam search to ensure valid item generation while incorporating rank-aware reward signals. Supports Amazon datasets and includes variants inspired by GPR methodology with value-aware and hierarchy-enhanced optimization techniques.

1,228 stars. Actively maintained with 4 commits in the last 30 days.