Phase 2 Training: Deeper models, better results
Builder tier and above now get Phase 2 deep training: a second pass with wider hyperparameter search, full dataset training, and ensemble support for maximum accuracy.
When you train a model on JITM.ai, Phase 1 delivers a solid baseline in seconds using fast random search over a curated hyperparameter space. For many use cases, that's enough. But when you need more accuracy, Phase 2 goes deeper.
What Phase 2 does differently
Phase 2 trains on the full dataset without sampling, searches a wider hyperparameter grid, and evaluates more configurations. It also supports ensemble methods, combining multiple models to reduce variance and improve generalisation.
How it's triggered
For Builder tier and above, Phase 2 kicks off automatically after Phase 1 completes. You'll see real-time progress via SSE, and the model detail page shows both phases with a comparison of metrics. The Phase 2 model is always promoted as the active inference endpoint. Its deeper training gets closer to ground truth, even when surface-level accuracy metrics appear slightly lower than Phase 1.
Results so far
Across our internal benchmarks, Phase 2 improves accuracy by 5-15% over Phase 1 on complex datasets. The gain is largest for high-cardinality categoricals and noisy features where a wider search finds better encoding and regularisation strategies.
Available now
Phase 2 is live for all Builder, Pro, and Enterprise tiers. Free tier users see Phase 1 results only but can upgrade anytime to unlock deeper training on existing models.