GenRec: A Unified Generative Recommendation Toolbox

Modular • Configuration-Driven • Reproducible

---

GenRec is an open-source, end-to-end framework designed to standardize the Generative Recommendation (GenRec) workflow. It provides a reproducible pipeline covering Representation → Tokenization → Modeling → Training → Inference.

📘 arXiv Paper (coming soon)

🔥 Motivation

Generative Recommendation is shifting the paradigm from scoring/matching to generative modeling. However, the current ecosystem is highly fragmented:

• Inconsistent Tokenization: Diverse quantization methods (RQ-VAE, VQ-VAE, OPQ, RKMeans) make inputs incompatible.
• Varied Backbones: Architectures range from Encoder–Decoder (T5/BART) to Decoder-only LLMs (Llama/GPT).
• Disparate Pipelines: Training and inference strategies (Beam Search vs. Prefix-tree) vary significantly between implementations.

The Result: Models are difficult to compare, hard to extend, and often unreproducible.

🎯 Our Goal

GenRec solves this by providing a single, plug-and-play stack that unifies the entire lifecycle.

• 🧩 Fully Modular: Decoupled components for Tokenization, Backbones, and Inference.
• ⚙️ Config-Driven: Manage complex experiments with simple YAML configurations.
• 📊 Fair Comparison: Benchmarking SOTA models (TIGER, Letter, RPG, etc.) under the same setting.
• 🔬 Standardized SID Modeling: The first open-source standardization for Semantic ID-based recommendation.

🔧 Pipeline Overview

Raw Data
↓
Download + Preprocessing
↓
Embedding Generation (Text / Image / CF / VLM)
↓
Multimodal Fusion (optional)
↓
Quantization (RQ-VAE / OPQ / PQ / RKMeans)
↓
Generative Recommender (TIGER / RPG / LETTER / LLMs)
↓
Inference (Beam Search / Prefix-tree / Contrastive Rerank)

---

🧱 Capability Matrix

Dimension	Category	Supported Components	Status
Data	Datasets	Amazon, MovieLens	✓
	Input Formats	Raw IDs, Embeddings, Codebooks (SID)	✓
Representation	Text	Qwen, T5, OpenAI Embedding API	✓
	Vision	CLIP ViT	✓
	Collaborative	SASRec	✓
	Fusion	Concat, MLP Fusion	✓
Quantization	Residual Family	RQ-VAE, Residual KMeans, Residual-VQ	✓
	Product Family	OPQ, PQ	✓
	Other	VQ-VAE, Multi-Codebook (RPG-style)	✓
Backbone	Encoder–Decoder	TIGER-style architectures	✓
	Decoder-only LLM	GPT-2, Qwen, LLaMA	✓
	Retrieval-Hybrid	RPG-style architectures	✓
Training	Objectives	LM Loss, Contrastive Loss, Hybrid Loss	✓
	Paradigms	SFT, Alignment, Multi-stage Training	✓
Inference	Decoding	Greedy, Beam Search	✓
	Constraints	Prefix-Tree	✓

🚀 Quick Start

Requirements

• Python 3.10 (recommended)
• CUDA 11.8+ (for GPU acceleration)
• PyTorch, CUDA, and other dependencies will be installed automatically via requirements.txt

git clone https://github.com/yourname/GenRec
cd GenRec
pip install -r requirements.txt

1 Data Preprocessing

We provide a dedicated submodule for downloading, cleaning, and extracting embeddings (Text/Image/CF).

👉 See detailed tutorial:
GenRec-Factory Data Processing & Embedding Guide

2 Quantization

Convert dense embeddings into discrete Semantic IDs (SIDs)

cd quantization

python main.py \
  --model_name rqvae \
  --dataset_name Musical_Instruments \
  --embedding_modality text \
  --embedding_model text-embedding-3-large

3 Generative Recommendation Models

Train a generative recommender using the generated SIDs.

cd recommendation

python main.py \
  --model TIGER \
  --dataset Baby \
  --quant_method rqvae