ster vector curve
How many pairs a direction needs: one method fitted at every chosen layer on growing numbers of pairs, each fit scored on the same held-out pairs.
Usage#
ster vector curve --model <MODEL> --method <caa|pca|logistic> --holdout <FRACTION> --sizes <N,N,...> [--pairs <PAIRS>] [--layers <LAYERS>] [--revision <REVISION>] [--device <DEVICE>] [--chat-template <auto|off>] [--precision <f32|f16|bf16>]The question wisent's optimize-sample-size asked. The last --holdout fraction of the pair set, in file order and rounded to whole pairs, is held out exactly as `ster vector optimize` holds it out. --method is then fitted at every layer on the first N pairs of the rest for every N in --sizes, and every fit is scored on the same held-out pairs. Ster states no method, fraction or size. --layers takes all, a comma list or a half-open range, as everywhere; --pairs may be omitted to use the active imported set.
What it reports#
points— one entry per size and layer:size,layer,holdout_accuracy,holdout_margin.holdout—fraction,fit_pairsandholdout_pairs, the cut every point was scored on.enough— per layer,best_accuracy(the best held-out accuracy any size reached there) andsmallest_size(the smallest size that reached it).method,trait_name, andchat_template— what this run encoded in.
wisent's optimize-classification is vector optimize, which ranks the logistic probe beside caa and pca at every layer on the same kind of holdout; its optimize is vector optimize then vector curve; and its optimization-cache is the artifact or report each run writes.
Desktop request#
ster request vector/curve takes pairs, method, holdout, sizes and optional layers, with the shared model fields, chatTemplate and precision, and returns the same document. Ster Desktop runs it from the Curve workflow.
Refusals#
| Sentence | Cause |
|---|---|
a size of <N> pairs does not fit in the <FIT> pairs left to fit on after holding out <HELD> | A size past the pairs left after the holdout. |
the held-out fraction must be above zero and below one | --holdout outside that range. |
a held-out fraction of <F> over <N> contrastive pairs leaves <FIT> pairs to fit and <HELD> to rank on; both need at least one | A fraction that empties either side. |
error: invalid value '0' for '--sizes <SIZES>' | A size of no pairs; a usage error, exit 2. |
missing field \sizes\`` | The desktop request names no sizes; refused with status 2 before anything loads. |