croissant-validator

name	croissant-validator
description	Validates and generates Croissant metadata for ML datasets. Use when checking dataset metadata compliance, creating new croissant.json files, validating existing metadata, or ensuring ML dataset descriptions follow the MLCommons Croissant format specification.
allowed-tools	Read, Bash, Write, Glob, Grep

Croissant Metadata Validator & Generator

Validate and generate croissant.json metadata for ML datasets.

Use This Skill When

• User asks to validate Croissant metadata.
• User asks to create or update croissant.json.
• User asks for MLCommons Croissant conformance checks.

Skill Layout

• scripts/validate_croissant.py: required-field checks + optional mlcroissant validation.
• scripts/generate_croissant_from_csv.py: generate metadata from CSV headers and sample values.

Keep executable code in scripts/; keep SKILL.md focused on workflow.

Python Environment Rule

For any command requiring package installs, use a temporary venv in the working directory and remove it after execution.

python3 -m venv .venv_croissant && \
.venv_croissant/bin/pip install -q mlcroissant pandas && \
# ... run command(s) ... && \
rm -rf .venv_croissant

Never run global pip install.

Validation Workflow

1. Locate metadata file (croissant.json, metadata.json, or user-provided path).
2. Run script-level validation:

   python3 skills/croissant-validator/scripts/validate_croissant.py path/to/croissant.json
   

3. If mlcroissant validation is required, run in venv:

   python3 -m venv .venv_croissant && \
   .venv_croissant/bin/pip install -q mlcroissant && \
   .venv_croissant/bin/python skills/croissant-validator/scripts/validate_croissant.py path/to/croissant.json && \
   rm -rf .venv_croissant
   

4. Report: pass/fail, missing required fields, structural warnings, and fix suggestions.

Generation Workflow (CSV)

1. Collect required inputs:
   • dataset name
   • description
   • SPDX license URL
   • creator name
   • dataset URL
2. Generate metadata from CSV:

   python3 skills/croissant-validator/scripts/generate_croissant_from_csv.py \
     data.csv \
     "My Dataset" \
     "Dataset description" \
     "https://spdx.org/licenses/MIT.html" \
     "Creator Name" \
     "https://example.com/dataset" \
     --output croissant.json \
     --cite-as "@misc{my-dataset-2026, title={My Dataset}, author={Creator Name}, year={2026}}"
   

3. Validate generated file with validate_croissant.py.
4. If needed, rerun generation or patch metadata manually.

Required Metadata Fields

• @context
• @type = sc:Dataset
• conformsTo = http://mlcommons.org/croissant/1.0
• name
• description
• license
• url
• creator
• datePublished
• distribution

High-Value Checks

• All @id values are unique.
• Each cr:FileObject has sha256 or md5.
• recordSet/field references are valid.
• Dates use ISO format (YYYY-MM-DD).
• License uses SPDX URL.

Common Fixes

• Add missing conformsTo.
• Replace non-SPDX license values with SPDX URLs.
• Add checksums to cr:FileObject entries.
• Add missing field.source mappings.

Resources

• Spec: https://docs.mlcommons.org/croissant/docs/croissant-spec.html
• GitHub: https://github.com/mlcommons/croissant
• Release: http://mlcommons.org/croissant/1.0