The language-mix ceiling, fixed on OpenStack before Wikimedia is measured: its own halves split docs from code¶
scripts/split_criteria.py checks an organization's registered split against OpenStack's, each
half deduplicated and split on its own as the runner reads a placebo half (it reproduces the
runner's 2,158 and 2,182 training examples on corpus v2). Generated from
split-criteria-openstack.json, OpenStack against itself:
| half | projects | train examples | largest project | its share | top suffixes |
|---|---|---|---|---|---|
| openstack-a | 119 | 2,158 | starlingx/docs | 0.208 | .rst 871, .py 692, .yaml 237 |
| openstack-b | 123 | 2,182 | starlingx/test | 0.148 | .py 1,362, .rst 313, .yaml 231 |
The halves' suffix mix differs by a total variation of 0.357, which is now the ceiling a Wikimedia or Chromium split must not exceed, and 2,158 the size floor. The ceiling is fixed here, before any Wikimedia split is computed.
What this says about H1 itself. H1's rationale, that both halves of one organization share its language mix, does not hold closely for OpenStack: one half is 40% reStructuredText, the other 62% Python. OpenStack's half-split contrast therefore carries a document-type difference alongside the project boundary, the confound the separability probe already named (a document type separates as well as an organization). The criterion only keeps a new organization from being worse on this than OpenStack. Whether H1 should carry a supplementary estimand restricted to a matched file type, as H2 carries one restricted to C++, is recorded as open and decided before Stage 1.