README.md

metadata

dataset_info:
  features:
    - name: instruction
      dtype: string
    - name: code
      dtype: string
    - name: response
      dtype: string
    - name: file
      dtype: string
  splits:
    - name: train
      num_bytes: 247832934
      num_examples: 16440
  download_size: 86431840
  dataset_size: 247832934
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - code
pretty_name: python docstring dataset
size_categories:
  - 10K<n<100K

Python Docstring Diff Dataset

This dataset contains training samples for models that generate Python documentation patches. Each example provides a Python source file with its docstrings removed and a corresponding unified diff patch that restores the documentation.

The dataset is designed for training or evaluating language models that assist with:

• Automatic code documentation
• Docstring generation
• Code review automation
• Developer tooling
• Dataset Structure

Each entry contains the following fields:

Field Description

instruction| Task instruction given to the model code| Python source code with docstrings removed response| A unified diff patch that adds the correct docstrings file| Original file path from the source project

Task Format

The model receives a Python file missing its documentation and must produce a unified diff that adds appropriate docstrings.

Example input:

def load_json(path):
    with open(path) as f:
        return json.load(f)

Example expected output:

--- a/file.py
+++ b/file.py
@@
 def load_json(path):
+    """Load JSON data from a file path."""
     with open(path) as f:
         return json.load(f)

Data Sources

The dataset was generated by scanning Python packages in github. Docstrings were extracted from functions, classes, async functions, methods, and modules using Python's AST parser. Low-quality documentation was filtered out using heuristics such as:

• Minimum docstring length
• Removal of TODO or placeholder documentation
• Deduplication of similar examples

Intended Use

This dataset is useful for training models that perform:

• automatic docstring generation
• documentation patch creation
• codebase documentation improvement tools
• AI-assisted code review systems

License

This dataset is released under the MIT License.