mlvault

Immutable, encrypted snapshots of ML training runs.

mlvault captures your model checkpoints, metrics, and logs after every training run — encrypted and stored in the cloud. Restore any run on any machine with a single command.

pip install mlvault mlvault-mlflow
mlvault init

Quick start

1. Install

pip install mlvault mlvault-mlflow

2. Configure your API key

Get a key at obsideo.io/mlvault-beta, then:

mlvault init
# Paste your mlvault API key: mlv1_...
# Validating key... ok
# You're all set.

3. Run and archive a training job

mlvault run --collect ./outputs python train.py

mlvault runs your script, then encrypts and uploads everything in ./outputs to cloud storage.

Run ID : myproject_20240307_143022_abc123
Collect: /home/user/project/outputs
Command: python train.py

-- Training started --
...
-- Collecting artifacts --
  3 artifact(s) -> myproject_20240307_143022_abc123.tar.gz
-- Encrypting --
  myproject_20240307_143022_abc123.tar.gz.enc (258.4 KB)
-- Uploading --

Done. Run archived: myproject_20240307_143022_abc123
Remote : myproject_20240307_143022_abc123_bundle.enc

4. Restore a run

mlvault restore myproject_20240307_143022_abc123

Works on any machine where mlvault init has been run with the same API key.

5. List runs

mlvault list

MLflow integration

mlvault integrates with MLflow as a custom artifact store. Artifacts logged via mlf.log_artifact() are staged locally, then pushed to cloud storage with mlvault commit.

Setup

import mlflow

# Set the artifact store once when creating an experiment
mlflow.create_experiment("my-experiment", artifact_location="mlvault://my-project")
mlflow.set_experiment("my-experiment")

Or set it via environment variable before training:

export MLFLOW_ARTIFACT_URI=mlvault://my-project

Training loop

import mlflow

with mlflow.start_run() as run:
    mlflow.log_param("lr", 0.001)
    mlflow.log_metric("loss", 0.42)
    mlflow.log_artifact("checkpoint.pt")

    run_id = run.info.run_id

# After training, push artifacts to cloud storage:
# mlvault commit <run_id>
print(f"Run ID: {run_id}")

Commit artifacts

mlvault commit <mlflow_run_id>

Restore MLflow artifacts

mlvault restore <run_id>

Commands

Security

Artifacts are encrypted with AES-256-GCM before leaving your machine. Encryption keys are derived from your API key — only you can decrypt your artifacts.

License

MIT