Forge-T5-Base-s1

Forge-T5-Base-s1

This model is initialized from the pre-trained google-t5/t5-base and fine-tuned on the AL-GR/AL-GR-v1 dataset using the FORGE framework for 4 training epochs.

More details can be found in the paper FORGE: Forming Semantic Identifiers for Generative Retrieval in Industrial Datasets.

Official Code: GitHub Repository

Evaluation Results on AL-GR/AL-GR-v1

Note: HR@K denotes Hit Rate at K — the proportion of test queries for which the correct answer appears in the top-K retrieved/generated results.

Usage

1. Download the Model

You can download this model locally using the huggingface_hub library:

import os
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"  # Optional: use mirror for faster download in some regions
os.environ["KMP_DUPLICATE_LIB_OK"] = "True"

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id='AL-GR/Forge-T5-Base-s1',
    local_dir='{YOUR_LOCAL_DIR}',  # Replace with your desired local path
    local_dir_use_symlinks=False,
)

2. Update Configuration

After downloading, update the configuration file used by the FORGE framework. Specifically, replace the load_checkpoint_from field in the JSON config file:

File: algr/config/generate_t5base_3layer_tiny.json

Update to:

"load_checkpoint_from": "{YOUR_LOCAL_DIR}"

Replace {YOUR_LOCAL_DIR} with the actual local path where you downloaded the model.

---

For more details about the training setup, dataset, or evaluation protocol, please refer to the FORGE framework repository.

Citation

If you find this work helpful, please cite the following paper:

@article{fu2025forge,
  title={FORGE: Forming Semantic Identifiers for Generative Retrieval in Industrial Datasets},
  author={Fu, Kairui and Zhang, Tao and Xiao, Shuwen and Wang, Ziyang and Zhang, Xinming and Zhang, Chenchi and Yan, Yuliang and Zheng, Junjun and others},
  journal={arXiv preprint arXiv:2509.20904},
  year={2025}
}