This repository contains the implementation of our paper:

RecBase: Generative Foundation Model Pretraining for Zero-Shot Recommendation.

git clone https://github.com/reczoo/RecBase.git cd RecBase pip install -r requirements.txt 

RecBase/ ├── data_process/ # data & embedding pipeline │ ├── amazon18_data_process.py │ ├── amazon_text_emb.py │ └── utils.py ├── models/ # CL-VAE model & trainer │ ├── clvae.py │ ├── layers.py │ └── ... ├── train.sh # one-click training ├── README.md └── requirements.txt 

# (1) ID + text cleaning + k-core filtering + chronological train/valid/test split python data_process/amazon18_data_process.py \ --dataset Games \ --input_path /raw/amazon2018 \ --output_path data/Games \ --user_k 5 --item_k 5 # (2) LLM text embedding (LLaMA-2 by default) python data_process/amazon_text_emb.py \ --dataset Games \ --root data/Games \ --plm_checkpoint meta-llama/Llama-2-7b-chat-hf \ --gpu_id 0

After finishing you will obtain
data/Games/Games.emb-llama2-td.npy (item text vectors) and all atomic files RecBole / SR-GNN ready.

Edit scripts/train.sh:

--data_path data/path_to_emb-llama2-td.npy

Then run: The script automatically loads embeddings → trains CL-VAE → saves latent codes & codebook.
Check-points are stored in ckpts by default.

Two files are ready after training:

The trained uniform index codebook will be used to further pre-train the dedicated llm.

If you use RecBase, please cite our paper:

@misc{zhou2025recbasegenerativefoundationmodel, title={RecBase: Generative Foundation Model Pretraining for Zero-Shot Recommendation}, author={Sashuai Zhou and Weinan Gan and Qijiong Liu and Ke Lei and Jieming Zhu and Hai Huang and Yan Xia and Ruiming Tang and Zhenhua Dong and Zhou Zhao}, year={2025}, eprint={2509.03131}, archivePrefix={arXiv}, primaryClass={cs.IR}, url={https://arxiv.org/abs/2509.03131}, }

Issues, PRs, new datasets & models are welcome! Let's make text semantics a first-class citizen in recommendation.