Changelog: extraction config

CHANGELOG.md

@@ -4,6 +4,11 @@ Changes that affect published data, newest first: one entry per publish, `## <da
 concise bullets on what changed in the data. `tasksource-jev-typed-decisions` releases are tagged on the Hub
 (`load_dataset("tasksource/tasksource-jev-typed-decisions", revision="v2026.10.13")`).
 
+## tasksource-jev-typed-decisions v2026.10.20.7
+
+- `extraction`: add web pages (WebSRC answer spans, SWDE attribute values in eight verticals), English span structure (Universal Dependencies EWT and GUM subtrees by relation, PreCo coreference mentions, Re-DocRED entities, CoNLL-2000 chunks), entity sources (BC4CHEMD chemicals, HarveyNER, BioNLP-CG, BioInfer, UNER English) and four procedural generators (typed mentions, business documents, nested configurations, raw HTML). 105 sources, 498,522 training, 35,743 validation and 38,574 test rows.
+- Sources of one collection share a training budget (20,000 rows; 80,000 for procedural generators), so suites split into many configs no longer dominate. Toxic spans and EBM-NLP outcomes are capped at 2,000 training rows after a label audit.
+
 ## tasksource-jev-typed-decisions v2026.10.20.6
 
 - `extraction`: add TAB personal-information masking (court cases; every mention that must be masked to anonymise a person, by type, from annotator majority) and WDC-PAVE product attributes (human-verified values or null per attribute of the category); 82 sources, 388,909 training, 27,720 validation and 30,257 test rows.