Split dataset

Performs a deterministic, in-place split of the project's dataset into "training", "testing", and optional "validation" sets. Split balancing can use the label, one or more metadata keys, or both as a composite grouping signal. Related samples can also be kept together across splits by metadata key. This is a deterministic process based on the hash of the name of the data. Returns immediately on small datasets, or starts a job on larger datasets. For example: { "trainingSplitRatio": 0.8, "testingSplitRatio": 0.1, "validationSplitRatio": 0.1, "excludeDisabledSamples": false, "stratifyBy": { "label": true, "metadataKeys": ["site", "scanner"] }, "keepTogetherMetadataKeys": ["capture_group"] } With these options, label/site/scanner are used to balance the split, while samples sharing the same capture_group value stay in the same split bucket.