Official implementation of the paper "Rethinking Tabular Data Understanding with Large Language Models" (https://arxiv.org/abs/2312.16702).

Start by cloning the repository to your local machine:

git clone https://github.com/Leolty/tablellm.git
cd tablellm

Create and activate a new environment, and install the required packages:

conda create -n tablellm python=3.10
conda activate tablellm
pip install -r requirements.txt

Unzip the dataset provided in the repository:

unzip assets/data.zip

After unzipping, you should have the following files:

└── data
    ├── wtq.json
    ├── tabfact.json

For replicating our study's findings, navigate to the scripts folder:

• scripts/all_dp.sh: Runs direct prompting on all wtq datasets.
• scripts/all_pyagent.sh: Runs python shell agent on all wtq datasets.
• scripts/vicuna_example.sh: An example of changing base model to vicuna on the subsampled wtq dataset.

  Ensure vllm is installed beforehand.

• scripts/perturbed_example.sh: An example of running experiments on perturbed wtq dataset.

Detailed explanations of parameters can be found in run_cot.py and run_agent.py.

For hands-on experience with the table agent, refer to the following notebook:

• examples/pyagent.ipynb

To access our experimental results on all wtq datasets, unzip this file:

unzip assets/results.zip

Post-unzipping, you will find:

└── results
    └── results/wtq-cot-all
        ├── result_5.json  # 5 runs of direct prompting, stored in a list
    └── results/wtq-agent-all
        ├── result_sc1.json
        ├── result_sc2.json
        ├── result_sc3.json
        ├── result_sc4.json
        ├── result_sc5.json # 5 runs of python shell agent, each in a separate file

For result evaluation, refer to the following notebook:

• examples/evaluate.ipynb

If you find this research useful in your work, please consider citing:

@misc{liu2023rethinking,
      title={Rethinking Tabular Data Understanding with Large Language Models}, 
      author={Tianyang Liu and Fei Wang and Muhao Chen},
      year={2023},
      eprint={2312.16702},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}