ster decisions split
Split labelled decisions into a training set and a held-out set, example by example, after a seeded shuffle.
Invocation and purpose#
ster decisions split --examples <EXAMPLES> --train-output <OUTPUT> --holdout-output <OUTPUT> --holdout <FRACTION> --seed <SEED>Examples are the unit, never questions: every question of one example lands on the same side, so a held-out state is one the model never saw under any question. Both sides are non-empty whenever the set has two or more examples; a fraction that would empty one side is clamped to one example. The same seed over the same set writes the same two files.
Inputs and options#
| Input | Required | Meaning |
|---|---|---|
--examples <EXAMPLES> | yes | The labelled decisions to split. |
--holdout <FRACTION> | yes | Fraction of examples held out, above zero and below one. Ster assumes none. |
--seed <SEED> | yes | Shuffle seed. |
--train-output <OUTPUT> | yes | Where the training examples are written. |
--holdout-output <OUTPUT> | yes | Where the held-out examples are written. |
Output and state#
Prints train and holdout, each with its path and examples count, and the seed used. Both files keep the input's example order within their side.
Refusals#
- Leaving out
--holdoutor--seedis a usage error that names the flag, exit 2. the held-out fraction must be above zero and below one.failed to read <path>,invalid calibration examples <path>and the labelled-decision validation refusals of `ster calibrate` when the input does not validate.