/rllm

Pytorch Library for Relational Table Learning with LLMs.

Primary LanguagePythonMIT LicenseMIT

rLLM (Documentation|Paper)

rLLM (relationLLM) is an easy-to-use Pytorch library for Relational Table Learning (RTL) with LLMs, by performing two key functions:

  1. Breaks down state-of-the-art GNNs, LLMs, and TNNs as standardized modules.
  2. Facilitates novel model building in a "combine, align, and co-train" way using these modules.

How to Try:

Let's run an RTL-type method BRIDGE as an example:

# cd ./examples
# set parameters if necessary

python bridge/bridge_tml1m.py
python bridge/bridge_tlf2k.py
python bridge/bridge_tacm12k.py

Highlight Features:

  • LLM-friendly: Modular interface designed for LLM-oriented applications, integrating smoothly with LangChain and Hugging Face transformers.
  • One-Fit-All Potential: Processes various graphs (like Social/Citation/E-commerce Networks) by treating them as multiple tables linked by foreigner keys.
  • Novel Datasets: Introduces three new relational table datasets useful for RTL model design. Includes the standard classification task, with examples.
  • Community Support: Maintained by students and teachers from Shanghai Jiao Tong University and Tsinghua University. Supports the SJTU undergraduate course "Content Understanding (NIS4301)" and the graduate course "Social Network Analysis (NIS8023)".

Todo List:

  • Code structure optimization
  • Large-scale RTL training
  • LLM prompt optimization
  • Support for more TNNs

Citation

@article{rllm2024,
      title={rLLM: Relational Table Learning with LLMs}, 
      author={Weichen Li and Xiaotong Huang and Jianwu Zheng and Zheng Wang and Chaokun Wang and Li Pan and Jianhua Li},
      year={2024},
      eprint={2407.20157},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2407.20157}, 
}