On en.wikipedia, many articles — especially well-developed ones — may use the "sfn" template or similar, in place of a ref tag. This does not contribute to the ref_tags count, even though it is used to replace a ref tag and the end result for a reader is essentially the same as using a ref tag.
For example, this article with more than 100 footnotes has only 4 ref tags:
Relatedly, the 'paragraphs without refs total length' covers nearly the whole article content, even though a majority of the prose is in paragraphs with references.
ORES models would likely be more accurate in quality predictions if ref-tag-equivalent templates were factored in. This would also improve the utility of the extracted features for downstream use (ie, it would help me out with outreachdashboard.wmflabs.org which now uses ORES features to show how many references get added/removed in any given revision).