How to Validate Data Before Training in GitHub Actions

Run schema, null, and range checks as a gate before training so a broken dataset fails CI instead of producing a broken model.

Add a validation step that checks the dataset schema, null rates, and value ranges (with pandas assertions or a tool like Great Expectations). Make it a dependency of the training job so training never starts on invalid data.

Steps

• Load the dataset and assert the expected columns and dtypes.
• Check null rates and value ranges against thresholds.
• Gate the training job on the validation job with needs:.

Workflow

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: pip install pandas
      - name: Validate dataset
        run: |
          python - <<'PY'
          import pandas as pd
          df = pd.read_parquet("data/train.parquet")
          assert set(["text", "label"]).issubset(df.columns)
          assert df["label"].notna().mean() == 1.0
          assert df["label"].between(0, 4).all()
          PY
  train:
    needs: validate
    runs-on: [self-hosted, gpu]
    steps:
      - run: python train.py --max-steps 1

Gotchas

• Validate before any GPU step so bad data fails on a cheap CPU runner.
• Version the expected schema with the data so the checks evolve together.

Related guides