Skip to content
Sushi ML Studio AI Platform & Data · v1.1.0 Launch on AWS

Release notes

What each version of Sushi ML Studio includes. To move to a newer version, launch a new instance from it; see the FAQ.

1.2.0

Released October 2026. Ubuntu 24.04, Python 3.12, scikit-learn 1.9.

Secure by default

  • The Studio is served over HTTPS at https://<public-ip>, with a trusted certificate for your instance's address that renews by itself. Ports 80 and 443 must be open in the security group. http://<public-ip>:8501 still works, and the sidebar points you to the encrypted address.
  • sushi-ml-studio https status shows how HTTPS is set up, and sudo sushi-ml-studio https on or off changes it. See Reach the Studio privately.
  • A password protects the Studio, starting as the instance ID, which you can replace with your own. See Choose your own password.

Large files

  • Before training, a model that may need more memory than the instance will have free, or take more than five minutes, says so on its card, with a sample size that should be fine.
  • Rows to train on, in a model's Adjust settings, trains it on a sample of the rows. It is scored on the same held-back rows as the others, so Results shows what the sample cost.
  • A model that runs out of memory stops on its own. Your other models and the rest of your work are kept, and Results offers to retrain it on a sample in one click.
  • Nearest neighbours and random forest finish several times sooner on large files, with the same models and scores.

Training and results

  • Seven models trained and compared for you, with ties called out and the simplest suggested. See Train your first model.
  • Adjust settings lets you choose a model's settings yourself and retrain just that model, judged against its defaults on the same rows.
  • Search for better settings tries several combinations and says whether it helped.
  • Random forest and extra trees work when the column to predict holds numbers stored as text, as Excel often does.

Export

  • Download one model, or every model as a single zip. See Predict and export.
  • File names say what the model is and when it was trained, and stay the same however many times you download.
  • The training code reproduces each model's score, including a model trained on a sample.

Last updated October 14, 2026 · Suggest an improvement