Transparency
Honest bias audit
We publish what the corpus over- and under-represents, so downstream models inherit known blind spots instead of hidden ones.
Preview · illustrative figuresWell represented
~78% recordsConcrete & reinforced structuresDBN V.2.6-98 family
92%Normative formal registerДБН / ДСТУ text
88%Central dialect regionKyiv, Cherkasy, Vinnytsia
81%Material terminologyNER label: material
76%Structural engineeringLoad-bearing, foundations
72%Under-represented
~22% recordsWestern dialect regionZakarpattia, Halychyna, Bukovyna
28%Field speech registerOn-site informal
24%Female speakersSpeech subset gender balance
18%Restoration / heritageDomain underrepresented
14%Southern dialect regionOdesa, Mykolaiv, Kherson
11%How we measure this. Strength ratio = share of corpus where a label, register or region is present vs. the balanced expectation (uniform across 27 regions, uniform across 9 NER types, 50/50 gender split). We publish the underweighted categories up front.
How we close the gap. We plan to prioritize under-represented categories in the annotation queue and, for speech, to recruit contributors from under-represented regions through the Contribute pathways.
