Official implementation of the paper "ALTER: Augmentation for Large-Table-Based Reasoning" . We warmly welcome discussions on table reasoning issues together!
ALTER: Augmentation for Large-Table-Based Reasoning
git clone https://anonymous.4open.science/r/tabular_data-295C
conda create -n alter python=3.10
conda activate alter
pip install -r requirements.txt
You can run the experiments with LocalFile Store or Redis Store.
To run the experiments with Redis store, you can use the following command to start a Redis store in Docker:
docker run -d -p 6379:6379 -p 8001:8001 redis/redis-stack:latest
# modif config.yaml first
sh scripts/aug.sh
sh scripts/pipelines.sh
We use the following datasets for the experiments:
-
scripts: This directory contains scripts for the main experiments.
-
config.yaml: This file contains the configs setting for the main experiments.
-
data_loader: This directory consists of python scripts to create PyTorch Dataset for different datasets and preprocessing utilities used for experiments
-
notebooks: This directory contains python notebooks that were used carry out the experiment in one run on benchmark datasets.
-
prompt_manager: This directory mainly contains the prompt template to conduct experiments
-
utils: This directory contains the code for all the util tools code for ALTER workflow, such as normalization or parsing.
-
batch_pipe.py: This file is the main logic for running the ALTER workflow.
-
run.py: start-up file for pipeline or augmentation process.
File Tree:
|-- ./augmentation.py # script for augmentation in pre-stage
|-- ./batch_pipe.py
|-- ./data_loader/__init__.py
|-- ./data_loader/datasets
|-- ./data_loader/table_augmentation.py
|-- ./data_loader/table_format.py
|-- ./data_loader/TableLoader.py
|-- ./executor/executor.py
|-- ./run.py
`-- ./utils.py
|-- notebook