https://github.com/andreaschandra/code-gen-extended

https://github.com/andreaschandra/code-gen-extended

Last synced: 15 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/andreaschandra/code-gen-extended
• Owner: andreaschandra
• Created: 2023-07-18T15:31:12.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-07-18T15:35:19.000Z (over 1 year ago)
• Last Synced: 2024-05-01T16:40:21.228Z (7 months ago)
• Language: Python
• Size: 7.64 MB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
### Steps for Reproducing Milestone 2 Results 

### Fine-tuning GPT-2 on CodeSearchNet for Code Generation

## Setup

From /cs230/gpt-2-csn:

$ pip install -r path/to/requirements.txt

$ python download_model.py 117M

## Data

- Note that CodeSearchNet python training data is stored in /cs230/gpt-2-csn/src/pythonTrain/python_train_all.txt and git lfs tracked for space efficiency.

- This dataset has been encoded and stored in /cs230/gpt-2-csn/src/pythonTrainPreprocessed/python_train_all.npz for space efficiency.

- To encode your own data, the following script is compatible with a minimum python 3.7.x: $ python encode.py trainingData.txt trainingData.npz

## Training

$ python /src/train.py --dataset /src/pythonTrainPreprocessed/python_train_all.npz --model_name 117M

- Samples will automatically be generated every 100 steps. 

- Additional parameters such as learning rate and batch size can be specified on the commandline.