build-number
duration-label: 3h 19m 11s
started-label-colon: time-hours-ago
Sweep the remaining tag penalties, errorth weight, and beam
Same train/held-out discipline as the +Prop sweep; all confirmed by
real builds and positive on both corpus halves plus the independent
hand-tagged eval.
- tags.reweight: every whole-class penalty was overpriced, as with
+Prop: +Der/Dimin 20->5, +Coll 15->4, bare hyphen 40->5,
+ABBR/+ACR/+Arab 20->5; +Foc/* is the exception and rises 13->26.
Combined +30 top-1, effects measured individually additive.
- ERRORTH_WEIGHT 40->32 (regenerated regex identical to
errorth-regen --weight 32): the reachability collapse sits below 30,
not below 40 as the old comment claimed; 32 over 30 on held-out
parity and 24 MB smaller. +3/+6/+8.
- config.json beam 28->80: top-1 is beam-invariant; 80 recovers 97%
of the 264-word offered-at-all gap (+135 top-5) at p95 254 ms,
within the parallel-execution latency budget.
Also audited, no change: final_strings.default.txt is load-bearing
(+25 top-1) but repricing fails held-out; words.default.txt all-help;
STRING_REGEX_EDIT_DISTANCE=2 is zero-gain at +17% size.
Production totals: top-1 9156 (85.77%), top-5 10371, offered 10544.
jobs-heading
divvun-actions cidivvun-actions ci
ran-in▶
Build Spellersdivvun-actions run lang-speller-build
ran-in▶
Build Grammar Checkersdivvun-actions run lang-grammar-build
ran-in▶
Build TTS Text Processordivvun-actions run lang-tts-textproc-build
ran-in▶
Test Spellersdivvun-actions run lang-speller-test
ran-in▶
Test Grammar Checkersdivvun-actions run lang-grammar-test
ran-in▶