Home

Awesome

JuICe Dataset

Code for the paper. This repository produces the dataset from the collected Jupyter notebooks. The dataset is available here if you don't want to run this pipeline. Modeling code is here.

Leaderboard

ModelDev BleuDev EMTest BleuTest EM
LSTM Baseline21.665.5720.925.71

Download notebooks

Get the notebooks here.

Setup

conda create -n {name} python=3.6
conda activate {name}

pip install -r requirements.txt

# decompress the downloaded notebooks file
unzip juice_notebooks.zip

Running pipelines

Produces the dataset. The pipeline requires around 254gb of disk space, and takes about 12 hours to complete on a 12 core machine.

./run_all {downloaded notebooks directory} {pipeline directory}

The datasets will be created under {pipeline directory}/final-dataset

Dataset Format

Each dataset record contains the following keys:

code_tokens_clean: The tokenized target code to generate.

context: A list of cells above starting with the cell directly above. If the cell is markdown the key nl will store the tokenized mardown. If it's a code cell the code_tokens_clean will store the tokenized code.