Skip to main content

Ship datasets

Collection isn't the product — the dataset is. This section covers the path from "annotated" to "released with a straight face":

  • Measure — inter-annotator agreement (Cohen's Îș, Gwet's AC1, weighted variants), per-annotator evaluation, quality dashboards.
  • Review — sampling-based review queues with accept / fix-accept / reject, escalations from automated QC and crowd validation surfaced first.
  • Export — every standard format, plus a first-class Hugging Face Hub integration: push fresh datasets with clean typed columns, enrich the source dataset you imported from, or fork it — with aggregation strategies (majority / adjudicated / first / raw), optional full-provenance sidecars, consent-gated speaker columns, ASR- ready 16 kHz audio derivatives, and speaker+prompt-disjoint splits (corpus engineering).

The result is a release your future reviewers can't poke holes in: license attested at collection, consent attached to every demographic column, provenance on every prompt, and no speaker or sentence leaking across splits.