Model Registry

Models need versioning too

'I trained a new model' is not enough information. Which model? What data? Which commit produced it? How did it score? A model registry answers these.

Common registries

Hugging Face Hub — public default; private orgs supported.
MLflow — open-source, self-hostable.
Weights & Biases (W&B) — full experiment tracking + registry.
SageMaker Model Registry — AWS-native.
Vertex AI Model Registry — GCP-native.

What metadata to record

Source SHA (Git commit that produced the model).
Training data version (DVC version, dataset SHA, S3 etag).
Hyperparameters.
Eval scores on the held-out set.
Train environment (Docker image SHA, GPU type, library versions).
Promotion stage (staging / production / archived).

CI integration

A successful training job uploads the model to the registry as 'staging'. A separate workflow (manual or scheduled) promotes to 'production' after eval gates pass. The serving stack pulls 'production' on startup.

Code

Push trained model to HF Hub on success·yaml

      - name: Train
        run: uv run python train.py --output ./out/
      - name: Eval
        id: eval
        run: |
          score=$(uv run python eval.py --model ./out/ --json | jq -r .accuracy)
          echo "score=$score" >> $GITHUB_OUTPUT
      - name: Push to HF Hub (staging only)
        if: steps.eval.outputs.score > 0.85
        env:
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          uv run python -c "
          from huggingface_hub import HfApi
          HfApi().upload_folder(
            folder_path='./out',
            repo_id='my-org/my-model',
            revision='staging',
            commit_message='train: ${{ github.sha }} score=${{ steps.eval.outputs.score }}',
          )
          "

Models need versioning too

Common registries

What metadata to record

CI integration

Code

External links

Exercise

Progress

Comments 0