wikalid / docs /NATIVE_REVIEW.md
jpaulpoliquit's picture
OurLID 1.0.0 (private preview): package, weights, model card, reproducible eval
863c520 verified
|
Raw History Blame Contribute Delete
1.19 kB

What needs native-speaker review before any release

  1. Gold set: re-label/adjudicate 299 docs / 593 spans (single non-native annotator now); priorities: hil (30), war (8), cbk (6), oth (Aklanon, Surigaonon etc.), nol boundaries.
  2. Tier-W training rows (web, GlotLID labels): audit 200 random rows per language, esp. hil, ceb, ilo, bcl, pam, pag; record label noise %.
  3. Seed marker lexicons (src/ourlid/seeds.py) and the induced closed-class lexicons (data/work/induced_lexicon.json): confirm top function words per language, flag Spanish/English loans and topic words.
  4. Frame vs majority convention for quoted English and the Taglish cs_level thresholds (0.03 / 0.25 English share): does the community agree with "light/heavy"?
  5. Benchmark: sanity-check 300 random items per category for label correctness (pam/pag/cbk Wikipedia sentences contain English/Spanish), natural-sounding SMS spellings, and add real social-media Taglish.
  6. Unsupported varieties: decide which of akl, sgd, mdh, cuyonon, sur etc. should become classes, and who can supply licensed text.
  7. Orthography: pre-1960s/old spellings, Chavacano orthographies, diacritics.