Overview
MLflow keeps a record of your machine learning work. You log the settings, results and files of each training run, then compare runs side by side in a web page. It also stores models and traces from AI applications.
Features
- Log parameters, metrics and files for every run, and compare runs in charts
- A model registry with versions and stages
- Traces for LLM and agent applications
- Prompt and evaluation tools for AI applications
- Artifacts such as models and datasets are stored on your box
Setting up
MLflow asks for a username and password after the box login. The username
is admin. The app makes a password for you. Open the app page in the
store and copy it:
- Username:
admin - Password: shown on the app page
Use the same username and password in your training code. Set
MLFLOW_TRACKING_URI to this app's address, and set MLFLOW_TRACKING_USERNAME
and MLFLOW_TRACKING_PASSWORD. Code on another machine can then log to this
server without the box login.
Good to know
The password on the app page is the first one. If you change it inside MLflow, the app page still shows the old one. Other users you add in MLflow can read everything by default. Give them edit permission on an experiment so they can log to it.
The first start downloads two small Python packages for the login, so it needs internet and takes a little longer. Later starts do not need it. MLflow's usage reporting to its makers is turned off.