TRAMS: Training-free Memory Selection for Long-range Langauge Modeling

📌 Table of Contents

Introduction
Development
Data
License
Citation
Contact

🚀 Introduction

This repository is for the EMNLP-2023 findings paper TRAMS: Training-free Memory Selection for Long-range Language Modeling.

💻 Development

Setup

Clone the repository from GitHub and install:

git clone https://github.com/lwaekfjlk/TRAMS.git
cd TRAMS/
pip install -r requirements.txt

Train Script

sh enwik8_xl_train.sh # for enwik8 XL train
sh wt103_xl_train.sh # for wt103 XL train

Test Script

sh enwik8_xl.sh # for enwik8 baseline
sh enwik8_trams.sh # for enwik8 trams
sh wt103_xl.sh # for wt103 baseline
sh wt103_trams.sh # for wt103 trams

Checkpoints

The Transformer-XL checkpoints for wikitext-103 and enwik8 are provided here.

📝 Data

We follow the instructions mentioned in Transformer-XL to collect enwik8 and wiki text-103 data. Data and its tokenized cache are provided here.

📜 License

This repository is released under the Apache-2.0 License.

📚 Citation

If you find this repository useful, please cite it as follows:

@misc{yu2023trams,
      title={TRAMS: Training-free Memory Selection for Long-range Language Modeling}, 
      author={Haofei Yu and Cunxiang wang and Yue Zhang and Wei Bi},
      year={2023},
      eprint={2310.15494},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

📮 Contact

If you have any questions or feedback, please feel free to reach out at haofeiy@cs.cmu.edu.