nb-clean cleans Jupyter notebooks of cell execution counts, metadata, outputs, and (optionally) empty cells, preparing them for committing to version control. It provides both a Git filter and pre-commit hook to automatically clean notebooks before they're staged, and can also be used with other version control systems, as a command line tool, and as a Python library. It can determine if a notebook is clean or not, which can be used as a check in your continuous integration pipelines.

Jupyter notebooks contain execution metadata that changes every time you run a cell, including execution counts, timestamps, and output data. When committed to version control, these elements create unnecessary diff noise, make meaningful code review difficult, and can accidentally expose sensitive information in cell outputs. By cleaning notebooks before committing, you preserve only the essential code and markdown content, leading to cleaner diffs, more focused reviews, and better collaboration.

For a detailed discussion of the challenges notebooks present for version control and collaborative development, see my PyCon UK 2017 talk and accompanying blog post.

nb-clean requires Python 3.9 or later. To run the latest release of nb-clean in an ephemeral virtual environment, use uv:

uvx nb-clean

To add nb-clean as a dependency to a Python project managed with uv, use:

uv add --dev nb-clean

Jupyter notebooks contain several types of metadata that nb-clean can handle:

Cell metadata includes information attached to individual cells, such as tags, slideshow settings, and execution timing. Cell metadata fields like collapsed, scrolled, deletable, and editable control notebook interface behaviour, whilst tags and custom fields support workflow automation.

Notebook metadata contains document-level information including the kernel specification, language version, and notebook format version. The language version information (metadata.language_info.version) frequently changes between Python versions and creates unnecessary version control noise.

Execution metadata encompasses execution counts for code cells and their outputs, along with execution timestamps and output data. This metadata changes every time you run cells, regardless of whether the actual code has changed.

You can check if a notebook is clean with:

nb-clean check notebook.ipynb

You can also process notebooks through standard input and output streams, which is useful for integrating with shell pipelines or processing notebooks without writing to disk:

nb-clean check < notebook.ipynb

When reading from standard input, nb-clean processes the notebook content directly without accessing the filesystem. This approach is particularly useful for automated workflows, continuous integration pipelines, or when you want to check notebooks without creating temporary files.

The check can be run with the following flags:

• To check for empty cells use --remove-empty-cells or the short form -e.
• To ignore cell metadata use --preserve-cell-metadata or the short form -m. This will ignore all metadata fields. You can also pass a list of fields to ignore with --preserve-cell-metadata field1 field2 or -m field1 field2. Note that when not passing a list of fields, either the -m or --preserve-cell-metadata flag must be passed after the notebook paths to process, or the notebook paths should be preceded with -- so they are not interpreted as metadata fields.
• To ignore cell outputs use --preserve-cell-outputs or the short form -o.
• To ignore cell execution counts use --preserve-execution-counts or the short form -c.
• To ignore language version notebook metadata use --preserve-notebook-metadata or the short form -n.
• To check the notebook does not contain any notebook metadata use --remove-all-notebook-metadata or the short form -M.

For example, to check if a notebook is clean whilst ignoring notebook metadata:

nb-clean check --preserve-notebook-metadata notebook.ipynb

To check if a notebook is clean whilst ignoring all cell metadata:

nb-clean check --preserve-cell-metadata -- notebook.ipynb

To check if a notebook is clean whilst ignoring only the tags cell metadata field:

nb-clean check --preserve-cell-metadata tags -- notebook.ipynb

nb-clean will exit with status code 0 if the notebook is clean, and status code 1 if it is not. nb-clean will also print details of cell execution counts, metadata, outputs, and empty cells it finds.

Note that the conflicting options --preserve-notebook-metadata and --remove-all-notebook-metadata cannot be used together, as they represent contradictory instructions.

You can clean a Jupyter notebook with:

nb-clean clean notebook.ipynb

This cleans the notebook in place. You can also pass the notebook content on standard input, in which case the cleaned notebook is written to standard output:

nb-clean clean < original.ipynb > cleaned.ipynb

The cleaning can be run with the following flags:

• To remove empty cells use --remove-empty-cells or the short form -e.
• To preserve cell metadata use --preserve-cell-metadata or the short form -m. This will preserve all metadata fields. You can also pass a list of fields to preserve with --preserve-cell-metadata field1 field2 or -m field1 field2. Note that when not passing a list of fields, either the -m or --preserve-cell-metadata flag must be passed after the notebook paths to process, or the notebook paths should be preceded with -- so they are not interpreted as metadata fields.
• To preserve cell outputs use --preserve-cell-outputs or the short form -o.
• To preserve cell execution counts use --preserve-execution-counts or the short form -c.
• To preserve notebook metadata (such as language version) use --preserve-notebook-metadata or the short form -n.
• To remove all notebook metadata use --remove-all-notebook-metadata or the short form -M.

For example, to clean a notebook whilst preserving notebook metadata:

nb-clean clean --preserve-notebook-metadata notebook.ipynb

To clean a notebook whilst preserving all cell metadata:

nb-clean clean --preserve-cell-metadata -- notebook.ipynb

To clean a notebook whilst preserving only the tags cell metadata field:

nb-clean clean --preserve-cell-metadata tags -- notebook.ipynb

Both the check and clean commands can operate on directories as well as individual notebook files. When you provide a directory path, nb-clean will recursively find all .ipynb files within that directory and process them. For example:

nb-clean check notebooks/

or

nb-clean clean experiments/

This is particularly useful for batch processing entire project directories or ensuring all notebooks in a repository are clean.

To add a filter to an existing Git repository to automatically clean notebooks when they're staged, run the following from the working tree:

nb-clean add-filter

This will configure a filter to remove cell execution counts, metadata, and outputs. The same flags as described above for interactive cleaning can be passed to customise the behaviour.

The Git filter operates by configuring the filter.nb-clean.clean setting in your repository's local Git configuration and adding the line *.ipynb filter=nb-clean to .git/info/attributes. This ensures that all notebook files are automatically processed through nb-clean when staged for commit. The filter configuration is local to the repository and won't affect your global or system Git settings.

To remove the filter, run:

nb-clean remove-filter

nb-clean can be used to clean notebooks tracked with Jujutsu rather than Git. Configure Jujutsu to use nb-clean as a fix tool by adding the following snippet to ~/.config/jj/config.toml:

[fix.tools.nb-clean]
command = ["nb-clean", "clean"]
patterns = ["glob:'**/*.ipynb'"]

The same flags as described above for interactive cleaning can be appended to the command array to customise the behaviour.

Tracked notebooks can then be cleaned by running jj fix. See the Jujutsu documentation for further details of how to invoke and configure fix tools.

nb-clean can also be used as a pre-commit hook. You may prefer this to the Git filter if your project already uses the pre-commit framework.

Note that the Git filter and pre-commit hook work differently, with different effects on your working directory. The pre-commit hook operates on the notebook on disk, cleaning the copy in your working directory. The Git filter cleans notebooks as they are added to the index, leaving the copy in your working directory dirty. This means cell outputs are still visible to you in your local Jupyter instance when using the Git filter, but not when using the pre-commit hook.

After installing pre-commit, add the nb-clean hook by adding the following snippet to .pre-commit-config.yaml in the root of your repository:

repos:
  - repo: https://github.com/srstevenson/nb-clean
    rev: 4.0.1
    hooks:
      - id: nb-clean

You can pass additional arguments to nb-clean with an args array. The following example shows how to preserve only two specific metadata fields. Note that, in the example, the final item -- in the arg list is mandatory. The option --preserve-cell-metadata may take an arbitrary number of field arguments, and the -- argument is needed to separate them from notebook filenames, which pre-commit will append to the list of arguments.

repos:
  - repo: https://github.com/srstevenson/nb-clean
    rev: 4.0.1
    hooks:
      - id: nb-clean
        args:
          - --remove-empty-cells
          - --preserve-cell-metadata
          - tags
          - slideshow
          - --

Run pre-commit install to ensure the hook is installed, and pre-commit autoupdate to update the hook to the latest release of nb-clean.

To ignore or preserve specifically the metadata defined in the nbformat documentation, use the following options: --preserve-cell-metadata collapsed scrolled deletable editable format name tags jupyter execution.

nb-clean can be used programmatically as a Python library, allowing integration into other tools.

import nbformat

import nb_clean

# Load a notebook
with open("notebook.ipynb") as f:
    notebook = nbformat.read(f, as_version=nbformat.NO_CONVERT)

# Check if the notebook is clean
is_clean = nb_clean.check_notebook(
    notebook, preserve_cell_outputs=True, filename="notebook.ipynb"
)

# Clean the notebook
cleaned_notebook = nb_clean.clean_notebook(
    notebook, remove_empty_cells=True, preserve_cell_metadata=["tags", "slideshow"]
)

The library functions accept the same configuration options as the command-line interface. The check_notebook() function returns a boolean indicating whether the notebook is clean, whilst clean_notebook() returns a cleaned copy of the notebook.

The following table maps from the command line interface of nb-clean 1.6.0 to that of nb-clean >=2.0.0.

The examples in the table use long flags, but short flags can also be used instead.

Copyright © Scott Stevenson.

nb-clean is distributed under the terms of the ISC license.