You pick the mode once, at the start of the run. Auto goes end to end unattended, taking the engine's own measured recommendation at every judgement call. Manual runs the identical engine and the identical measurements, but stops at the decisions where a human answer genuinely means something — which columns to remove or rename, what to do with each outlier column, which target to model, the fill strategy for every column with gaps — and hands each one to you, with its recommendation shown and marked as a recommendation. The measured contests further down the pipeline (which feature set wins, which projection wins, which balancing technique wins) are settled by cross-validation in both modes, and reported rather than prompted.

The fear worth naming: most data professionals have used a tool that asks for a decision and then quietly does something else. In OreML a manual step cannot continue with a single column left undecided, your answer has to be one of the options that step genuinely supports, and the value you chose is what gets applied, what the step's audit record stores, and what the steps downstream read. The live log narrates it as it happens and the export is on disk immediately — so you never have to take our word for it. Choose manual whenever domain knowledge should beat a statistic; choose auto when you want the statistic.

OreML can be driven with up to two concurrent targets on any hardware — but in manual mode, where you are the one choosing them, we strongly advise running one target at a time. Two targets double the per-step decisions you have to hold in your head, and every downstream stage forks in two. A single-target run gives you cleaner results, lower memory pressure, and output that's far easier to interpret. Reach for two targets only when you specifically need a side-by-side comparison.

In manual mode, you decide upfront whether a run is supervised (with a target column) or unsupervised (without one). OreML's built-in guards will catch an obviously wrong or contradictory choice, but this single decision reshapes the entire pipeline — target discovery, leakage checks, class balancing, and evaluation all follow from it. It's the one choice worth an extra moment's thought.

OreML is built to run on modest hardware and adapts automatically to your system and your data — but on larger datasets, memory is what you'll feel. The more RAM your machine has, the larger the datasets it handles comfortably. On a modest machine the pipeline still runs to completion; a big dataset simply takes longer to get there.

OreML doesn't make you wait for the very end: it writes analysis- and ML-ready output at six checkpoints along the run — after schema validation, after each of the two outlier stages, after feature unification, after dimensionality reduction, and after balancing — in standard tabular formats you can open in your own tools immediately. That is the default, and we recommend keeping it. You can switch to exporting after every single step, but it is genuinely verbose: far more files, far more disk, and the six that matter get buried among them. Reach for it when you are debugging a specific stage, not as a habit.