DANDI AI Notebooks

This repository automates the generation of Jupyter notebooks for analyzing DANDI datasets (dandisets). It processes multiple dandisets and creates customized analysis notebooks based on configuration files. The automation uses dandi-notebook-gen and minicline.

This packages were created as part of the Pre-COSYNE Brainhack, March 2025, Montreal.

View results (updates regularly)

View ratings (updates regularly)

Repository Structure

dandisets/
  ├── [dandiset_id]/
      └── [date]/
          ├── config.yaml       # Configuration for notebook generation
          ├── [dandiset_id].ipynb   # Generated analysis notebook
          ├── metadata.json     # Generation metadata and system info
          └── working/         # Working directory with analysis scripts
              ├── notebook.py
              ├── notebook.ipynb
              └── tmp_scripts/  # Generated analysis scripts and plots

How It Works

1. The script scans through directories in dandisets/, each representing a DANDI dataset
2. For each dandiset, it processes dated subdirectories containing:
   • A config.yaml file specifying the model to use
   • Generated analysis notebooks and scripts
3. For each configuration:
   • Creates a working directory for analysis files
   • Generates a Jupyter notebook using the specified model
   • Produces analysis scripts and visualizations
   • Records metadata including version info and system details

Configuration

Each dandiset directory requires a config.yaml file specifying:

• model: The model to use for notebook generation

Generated Files

For each processed dandiset:

• Main notebook file ([dandiset_id].ipynb)
• Metadata file with generation details
• Working directory containing:
  • Analysis scripts
  • Generated plots
  • Execution logs
  • Temporary processing files

Usage

Creating New Configurations

Use add_config.py to create new dandiset configurations:

python add_config.py <dandiset_id> <model>

This creates a new dated directory with a config.yaml file. For models with paths (e.g., 'deepseek/deepseek-r1'), only the final part is used in the directory name.

Grading notebooks

To grade a notebook using the notebook-grader package, first install the latest version of the package:

# Clone https://github.com/magland/notebook-grader
cd notebook-grader
git pull
pip install -e .

Then run the grading script:

cd dandisets/<DANDISET_ID>/<SUBDIR>
notebook-grader grade-notebook <DANDISET_ID>.ipynb --auto --output-notebook <DANDISET_ID>_graded.ipynb --output-json <DANDISET_ID>_graded.json --model google/gemini-2.0-flash-001

The output notebook will contain new markdown cells that identify potential problems, if any are found.

Requirements

• Python packages:
  • dandi_notebook_gen
  • pyyaml
  • jupyter

Notes

For the gh workflow to be able to push updates back to the repo you need to enable "Read and write permissions" in Settings -> Actions for the repo.