ml-validate 0.1.0

Offline-first ML dataset metadata validator for NFDI schemas (Croissant, ML-Schema, XDM, Chemotion, and more)

ml-validate

Offline-first Python validator for NFDI and ML metadata schemas.

Features

• Bundled schema registry (no network required for schemas)
• Python API via Validator and Profiler
• CLI for CI/CD pipelines
• Supports Croissant, ML-Schema, JSON Schema-based formats, and repository schemas

Install

pip install ml-validate

For URL validation support:

pip install "ml-validate[fetch]"

Quickstart

from ml_validate import Validator

validator = Validator()
result = validator.validate("dataset.json", "croissant-1.0")
print(result.passed, result.score)

ml-validate validate dataset.json --schema croissant-1.0

Example Data (Chemotion JSON-LD)

This repository now includes reusable example files in examples/ derived from:

• ../example_datasets/JSON-LD_Analysis_874350-20260219122553.json

Included examples:

• examples/chemotion_jsonld_sample.json: curated, realistic JSON-LD sample
• examples/chemotion_jsonld_incomplete.json: intentionally incomplete sample for failure demos

The sample keeps the metadata structure and key fields used by the validators while avoiding large personal-author blocks.

CLI Examples

Validate a realistic Chemotion-style JSON-LD example with Croissant heuristics:

ml-validate validate examples/chemotion_jsonld_sample.json --schema croissant-1.0

Run a profile across two schemas and get machine-readable output:

ml-validate profile examples/chemotion_jsonld_sample.json --schemas croissant-1.0,w3c-ml-2.0 --output json

Show a failing validation case:

ml-validate validate examples/chemotion_jsonld_incomplete.json --schema croissant-1.0

Python API Examples

from pathlib import Path
from ml_validate import Profiler, Validator

payload = Path("examples/chemotion_jsonld_sample.json").read_text(encoding="utf-8")

validator = Validator()
single = validator.validate(payload, "croissant-1.0", source_type="json")
print(single.passed, single.score)

profiler = Profiler()
results = profiler.profile(payload, schema_ids=["croissant-1.0", "w3c-ml-2.0"], source_type="json")
for schema_id, result in results.items():
    print(schema_id, result.status, result.score)

CI/CD Example (GitHub Actions)

You can validate metadata files automatically in pull requests and fail the build when rules are not met.

PR validation workflow

Create .github/workflows/metadata-validation.yml:

name: Metadata Validation

on:
  pull_request:
    paths:
      - "datasets/**/*.json"
      - ".github/workflows/metadata-validation.yml"

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
        uses: actions/checkout@v4

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Install validator
        run: pip install ml-validate

      - name: Validate a dataset (strict)
        run: |
          ml-validate validate datasets/metadata.json \
            --schema croissant-1.0 \
            --fail-under 90

Multi-schema profile gate

Use profiling if your pipeline must satisfy more than one schema:

- name: Profile against multiple schemas
  run: |
    ml-validate profile datasets/metadata.json \
      --schemas croissant-1.0,w3c-ml-2.0 \
      --fail-under 80

Publish workflow note

If you publish this package from GitHub Actions, see:

• .github/workflows/publish.yml

That workflow builds on tag push (for example v0.1.0) and uploads distributions to PyPI.