Calibrate

Recalibrates the starting confidence of each path on a team's own corpus. Issue bodies that are issue-form output (### Label sections) state their environment under labelled headings; those section facts are the truth, the sections are cut out, and the extractor reads what is left. A path's base is scaled by its accuracy over its mean confidence on the dev side of a seeded split, and the held-out side scores the result.

Properties

Link copied to clipboard
const val DEV_BELOW: Int = 128
Link copied to clipboard
const val MIN_FACTS: Long = 100
Link copied to clipboard
const val SEED: String

Functions

Link copied to clipboard

Equal values, or one cut at a dot from the other, as Combine does.

Link copied to clipboard
fun isDev(seed: String, repo: String, number: Long): Boolean

True on the dev side; the split depends only on seed, repo and number.

Link copied to clipboard
fun run(lines: Iterable<String>, seed: String = SEED, minFacts: Long = MIN_FACTS): Calibration

lines is a corpus file as buildCorpus writes it: an opening line, one JSON issue per line (after the first with a leading comma) and a closing ].

Link copied to clipboard
fun withoutSections(body: String, truth: List<Fact>): String

body without the heading blocks that hold a section fact in truth.