GenRec

A Model Zoo for Generative Recommendation.

Benchmark Results

Validation Set Metrics

Amazon 2014 Beauty

Methods	R@5	R@10	N@5	N@10
TIGER (Paper)	0.0454	0.0648	0.0321	0.0384
TIGER (Ours)	0.0465	0.0721	0.0297	0.0378

Amazon 2014 Sports

Methods	R@5	R@10	N@5	N@10
TIGER (Paper)	0.0264	0.0400	0.0181	0.0225
TIGER (Ours)	0.0266	0.0414	0.0176	0.0224

Amazon 2014 Toys

Methods	R@5	R@10	N@5	N@10
TIGER (Paper)	0.0521	0.0712	0.0371	0.0432
TIGER (Ours)	0.0420	0.0647	0.0280	0.0350

Installation

pip install -r requirements.txt

Supported Datasets

• Amazon Beauty
• Amazon Sports
• Amazon Toys

Data will be automatically downloaded on first run.

RQVAE Training

RQVAE generates semantic IDs for items using residual quantization.

# Train on Beauty (default)
python genrec/trainers/rqvae_trainer.py config/rqvae/amazon.gin

# Train on other datasets
python genrec/trainers/rqvae_trainer.py config/rqvae/amazon.gin --split sports
python genrec/trainers/rqvae_trainer.py config/rqvae/amazon.gin --split toys

TIGER Training

TIGER is a generative retrieval model for sequential recommendation.

# Train on Beauty (default)
python genrec/trainers/tiger_trainer.py config/tiger/amazon.gin

# Train on other datasets
python genrec/trainers/tiger_trainer.py config/tiger/amazon.gin --split sports
python genrec/trainers/tiger_trainer.py config/tiger/amazon.gin --split toys

Note: TIGER requires a pretrained RQVAE checkpoint. Train RQVAE first.

Command Line Options

Dataset Selection

--split <dataset>  # beauty, sports, toys

Gin Parameter Override

--gin "param=value"  # Override any config parameter

Examples

# Change epochs and batch size
python trainer.py config.gin --gin "train.epochs=200" --gin "train.batch_size=128"

# Set wandb run name
python trainer.py config.gin --gin 'train.wandb_run_name="my_experiment"'

Citation

If you find this project useful, please cite:

@software{genrec2025,
  title = {GenRec: A Model Zoo for Generative Recommendation},
  author = {Qi Lu},
  year = {2025},
  url = {https://github.com/phonism/genrec}
}

References

• TIGER: Recommender Systems with Generative Retrieval
• RQ-VAE-Recommender by Edoardo Botta