MSMARCO dataset is obtained from https://github.com/liyongqi67/MINDER.
code for offline hierarchical RQ-VAE training.
bash train_tokenizer.sh
bash tokenize.sh
bash train.sh
bash test.sh
This is a version that utilize pytorch DDP to train hierarchical RQ-VAE. The embeddings of all items can be divide into several npy files, named "semantic_emb_*.npy".
We directly use LLama-Factory to train our GR model based on Qwen2.5-1.5B-Instruct. LLama-Factory can be accessed from https://github.com/hiyouga/LLaMA-Factory.