blastwind's picture
Update README.md
d8fefa9 verified
metadata
dataset_info:
  features:
    - name: lang
      dtype: string
    - name: seed
      dtype: string
  splits:
    - name: train
      num_bytes: 3114466
      num_examples: 10000
  download_size: 1629429
  dataset_size: 3114466
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

This dataset contains 10000 random snippets of 5-15 lines parsed from bigcode/starcoderdata.

Specifically, I consider 10 languages: Haskell, Python, cpp, java, typescript, shell, csharp, rust, php, and swift. And, I collect 1000 documents for each language, and then extract 5-15 random lines from the document to create this dataset.

See MagiCoder and their seed collection process. In my usecase, I needed some inspiration documents for generating synthetic datasets.