Nebula: Code Generation through Natural Language Understanding

📑 Paper | 🌐 Project Page | 💾 Released Resources | 📦 Repo

We release the base training data for our Nebula code generation models, curated from the original CodeSearchNet corpus maintained by the GitHub research team.

The data format for each line in the train_code_v3.jsonl is as follows:

{
  "code_snippet": <the original code snippet>,
  "docstring": <the natural language documentation>,
  "language": <programming language identifier>,
  "func_name": <the function name>,
  "repo": <source repository name>,
  "path": <file path in original repository>,
  "metadata": <additional context information>
}

Some entries have truncated docstrings due to maximum length constraints during preprocessing.

*Note: We filtered samples based on code quality metrics. Future versions may include additional quality annotations.

License

The license for this dataset is MIT.