Skip to content

Type migration with large language models for code. Migrates JavaScript to TypeScript by predicting type annotations and generating type definitions.

Notifications You must be signed in to change notification settings

nuprl/StenoType

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

StenoType

Type migration with large language models for code. Migrates JavaScript to TypeScript by predicting type annotations and generating type definitions.

This is the code repository for the dissertation Predicting TypeScript Type Annotations and Definitions With Machine Learning, specifically, Chapter 5.

The training dataset is on Hugging Face. Parts of the code may refer to it as ts-training-get4. This is a preprocessed version of ts-training, revision v1.1p1.

The final StenoType model is on Hugging Face. You will need to accept the agreement to access the model. The code and results may refer to this model as stenotype-7b-a6d445d-ckpt1000, as it was fine-tuned based on commit a6d445d.

There are two evaluation datasets: stenotype-eval-ts (also called stenotype-eval-dataset-subset in the code and TS-Sourced in the dissertation) and stenotype-eval-js (also called typeweaver-bundle-filtered-subset in the code and JS-Sourced in the dissertation). To type check the stenotype-eval-js dataset, you will also need to download the tarball from Hugging Face.

Figures and result summaries are in the results/ directory. Full results are on Hugging Face.

Instructions

  1. Clone the repository:
git clone [email protected]:nuprl/StenoType.git
cd StenoType
git submodule update --init --recursive
  1. Follow the instructions to set up miniconda.

  2. Create a conda environment with Python 3.11 and install dependencies:

conda create -n gpu python=3.11
conda activate gpu
pip install -r requirements.txt
conda install -c conda-forge nodejs=20.8.1
npm install -g --no-save [email protected]
  1. Download the StarCoderBase-7b and StenoType models:

    a. Ensure that you have a Hugging Face account.

    b. Accept the agreements for StarCoderBase-7b and StenoType.

    c. On the command line, log into Hugging Face with huggingface-cli login.

    d. In a directory of your choosing, e.g. ../models, run git clone [email protected]:bigcode/starcoderbase-7b and git clone [email protected]:nuprl/stenotype.

    e. To save space, you can delete the .git directory (and possibly pytorch_model*.bin if model*.safetensors already exists).

  2. Accept the agreement for the ts-eval evaluation dataset.

  3. Now you can run the experiments:

# See what configurations can be wron
python src/main.py --show_configs

# To run inference on config 0 (this is very slow):
python src/main.py --infer 0

# To evaluate (this is CPU-bound):
python src/main.py --evaluate

# To generate dataset-level summaries (this is pretty fast):
python src/main.py --summarize
  1. To browse the results, you can use the viewer. Type "help" for help.
python src/viewer.py --dataset path/to/results/dataset.parquet

Dependencies

  • git
  • Python 3

Using Conda or virtual environments is recommended.

About

Type migration with large language models for code. Migrates JavaScript to TypeScript by predicting type annotations and generating type definitions.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published