Repository navigation
⚠️ Read this before you importIf you already have the 1.0 deck, delete all twenty 1.0 decks first — or
import into a fresh Anki profile. Importing 1.1 on top of 1.0 does not
work cleanly: the note type has a new field (Gloss_Source) and 70 notes
changed their part of speech, which changes their identity, so a plain
re-import leaves the old notes sitting beside the new ones.Deleting the 1.0 decks loses the review history on them — the scheduling,
the intervals, the lapses — and there is no way to carry it across. 1.0 was
never announced, so this should affect very few people.
Glosses. Every one of the 10,000 entries now carries an English meaning, and
9,808 of them a real sentence from the corpus with its translation, so the
Anki deck works in both directions from the first import instead of shipping
blank for you to fill in. A native-speaker Portuguese teacher checked the top
500 over two rounds and her corrections are in the published glosses; the
9,500 below that are the model's work alone, and the README says so.
Along the way the pipeline gave up several errors the glosses made visible: 51
rows that were never words, são counted as a noun rather than a million
occurrences of ser, and every contraction labelled by what it contracts.
Added
- English glosses and example sentences,
out/glosses.tsv: one to
three English senses per entry, commonest first, plus one sentence taken
verbatim from the corpus and its translation, and flags forvulgar,
bp-leaning,archaic,name-likeanduncertain. Written by Claude
(claude-opus-5), one request per entry, from the entry, its part of
speech, its frequency and up to twenty real corpus sentences containing
it. The prompt ships with the repository
(scripts/gloss_prompt.md) and the run is pinned inconfig.yaml.
These are a snapshot of one model's answers, not a reference work, and
are not reproducible byte for byte the way the lists are. - A repair step between the model's reply and the published row
(gloss.repair). Three things went wrong often enough in the first full
run to be worth handling rather than failing over, and each is handled in
a way that keeps the published example a real corpus line: a
doubly-escaped\uXXXXis decoded (5 rows); an example the model tidied
itself — dropped the dialogue dash — is re-anchored to the corpus line it
tidied, which changes nothing about the card (6 rows); and an example the
model edited (an accent added, a pronoun supplied, a clause trimmed) is
dropped, because there is no telling a correction from a corruption, and
the entry keeps its gloss alone (24 rows). Every one is counted and listed
in the gate report. - Gates on the glosses,
reports/gloss_gates.md: coverage, and the
check that matters — every example sentence must be one of the corpus
lines the model was shown, character for character, and must contain the
entry. A run failing any fatal gate is not published. Per-flag counts and
the cost of the run are in the same report. out/glosses.tsvis a release asset, attached to the tagged release
alongside the.apkgand the two list TSVs, so the glosses can be used
without cloning the repository.out/SHA256SUMS.txt, the SHA-256 of all four release assets in
sha256sum -cformat. The build writes it after the assets exist, so the
digests describe what the pipeline produced rather than what someone
pasted into the release notes, and a rebuild that moves an asset moves its
line. Which files count as assets isoutput.release_assetsin
config.yaml. Anyone pinning this data in another project should pin the
digest rather than the tag.eval/gloss_overrides.tsv, a reviewer's answer beating the model's,
in the same shape as the other human-decision files ineval/: match on
lemma + pos, and only the fields a row fills in are taken, so a row can
replace the gloss, the example, or both. An overridden row is published
withsource = overrideinout/glosses.tsvand is exempt from the
corpus-provenance gates — a reviewer's sentence has a different
provenance, not a broken one — but the gate counts it, and counts
separately any overridden example that is not one of the corpus lines the
model was shown. An override naming an unpublished lemma is an error, so a
typo cannot fail silently. Seeded with the one entry the model will not
gloss (caseiro).- A review sample,
eval/gloss_review.tsv: all of ranks 1–500 plus 200
spread evenly over the rest, with the sentences each entry was offered, so
a reader can check what the model was working from. - A targeted corpus pass for entries the sampled contexts missed. Nine
published entries — eight multi-word expressions and one folded adjective
— had no sampled sentence containing them, because pass 2 keeps lines for
the 70,000 most frequent forms and, for a phrase, only lines that happen
to contain the whole phrase. Those now get one scan of their own, choosing
lines by the same digest rule, so every one of the 10,000 entries was
glossed from real sentences rather than from its headword alone.
Known issues
- One gloss is not the model's.
caseiro(rank 4209, "homemade") trips
a safety classifier on every attempt, so its gloss is written by hand in
eval/gloss_overrides.tsvand the row is markedsource = override.
Every other gloss in the file is the model's. Thegloss_presentgate is
back to zero tolerance now that overrides cover refusals. - 200 entries have no example sentence (2.0%). Either every sampled line
showed a different word —domésticathe adjective rather than the noun —
or the lines were unintelligible fragments. Those entries have a gloss and
no example; the reverse Anki card still works, the sentence fields are
just empty. - The glosses are not reproducible byte for byte. The lists are; these
are one model's answers on one day, and the response cache incache/gloss/
is what makes a rerun repeat them rather than re-derive them.
Fixed
-
deverwas published as a noun at rank 68, carrying the verb's whole
2,220 per million. The commonest modal verb in the language — devo,
deve, devia, deveria — had no row of its own, and 1.0.0 had the
same error. The POS column decided verb-ness from the bare infinitive
alone, and a baredeverin subtitles is usually o dever, "the duty":
43 of its 50 sampled sentences are tagged NOUN, so the lemma was ruled
out as a verb and all 857 VERB tags across 36 inflected forms were
thrown away. Nothing survived but the noun.The POS column now reads the whole paradigm (
fixes.paradigm_verb_pos):
a lemma the infinitive test rejects still counts as a verb when its
VERB/AUX tags come from at leastpos.verb_paradigm_min_surfaces(4)
distinct inflected forms that share its stem. The stem rule is what keeps
mar— whose stem is the single letterm— from collectingmandoand
marqueand reading as a verb. The test can only ever add a verb
reading, so nothing loses one, and it is used for the POS column alone:
lemma resolution and the ser/ir split keep the old test and do not move.Four other entries gained a verb reading the same veto was hiding:
jantar(noun 651, verb 1372),colar(noun 3070, verb 3924),tremer
(noun 4027, verb 5626), anddorme, which wasintjand is nowverb.
deveritself does not split: the noun is 1.9% of its occurrences,
far belowpos.split_min_share, so the ~26,600 uses of o dever are
counted inside the verb row.colherkeeps one noun row for the same
reason, its verb reading missing the threshold at 24.5%. Three entries at
the foot of the list —presunção,bipolar,cicatrizar— were pushed
out of the 10,000 by the three new rows. No other count moved.Because the published part of speech is half of an Anki note's identity,
deveris a different note in 1.1.0 than in 1.0.0 — and a different entry
for the reviewer, whose verdict had been given on the noun. She signed the
new verb row off separately, so the top 500 is still reviewed in full; the
noun example she had written is held ineval/gloss_overrides.tsvas a
comment, to be restored if a noun row ever returns.jurarhas a related
but different fault and is left for a later release (#23). -
51 rows were not words at all:
conheçoa,deixaa,levaaand 48 more.
conheçoasat at rank 5,290 in 1.0.0, glossed "I know her". They were
conheço-a,deixa-a,leva-a— properly hyphenated in the corpus — and
the tokenizer glued them back together.aandasare both enclitic pronouns and future/conditional tense
infixes, andsplit_encliticreassembled a mesoclitic verb whenever the
tail it had peeled ended in one. Mesoclisis always puts the clitic
between stem and infix (dar-lhe-ia= daria + lhe), so a tail of one
can never be an infix; that single condition was missing. Every affected
row ended in-aor-as, which is what gave the cause away — genuine
hyphen loss would not favour one clitic so heavily.The tokens went where they belong: deixar +28,981, levar +19,546,
matar +11,196, conhecer +7,941. -
Clusters the corpus writes without their hyphen are now repaired
(fixes.repair_glued_enclitics):deixame,dáme,calate,sêloand
26 more, 1,556 tokens. This is the smaller half of the problem and it
cannot be done by rule — stripping a clitic-shaped tail and checking that
the stem is a verb form also breakssera(a missing accent on será),
saiste(saíste),rodeo,eramosandfodeste. So
scripts/make_glue_table.pygeneratesscripts/data/enclitic_clusters.tsv
from clusters the corpus writes both ways, and the tokenizer consults
the table rather than guessing. A cluster has to be attested hyphenated at
least 100 times, at least twice as often as glued, be neither a dictionary
word nor a recognised verb form, and not be capitalized mid-line — that
last test is the one that matters:mateooccurs 1,113 times away from the
start of a line and is capitalized every single time. It is Mateo, not
mate-o.nola,pirate,riseandriteare the same story. The
table's digest is hashed into every cache key, so regenerating it
invalidates the passes that depend on it.Effect on the list: 52 rows out, 52 in. 636,582,535 tokens (+368,083)
over 827,895 types (−10,463). The gold-set score is unchanged at 98.5%, the
quality gate still finds 0 suspects, and theser/irrule fingerprint is
unchanged (2defd01d6b1615a7, 95.9%), so the rule still applies: ser
21,334,148 and ir 9,023,260, both within 400 of before. One multi-word
entry (guarda costeira) dropped out as bigram counts shifted and
efeitos colateraistook its place; the other 51 replacements are new
entries at the bottom of the list, ranks 9,904–10,000. -
The Anki deck ships with the glosses. All 10,000 notes carry
Gloss_EN, andExample_PT/Example_ENwhere there is an example
(9,808 of them; the other 192 are entries no sampled sentence illustrated).
Every note therefore has an English → Portuguese card as well as a
Portuguese → English one: 20,000 cards, where 1.0.0 had 10,000. The package
is 5.9 MB, up from 3.5 MB.One field is added at the end of the note type.
Gloss_Sourcesays
where the row came from, in small print beside the rank and frequency —
"Claude; example taken verbatim from the corpus" for 9,972 of them, and for
the other 29 the wording that says a native speaker chose the sentence,
edited it, corrected the English, or that the gloss was settled by hand.The note type's field list is frozen from 1.1.0. Anki matches fields by
position, so adding, removing or reordering one breaks a re-import for
anyone already studying the deck. Anything further goes in a new note type.
The note type id, the twenty deck ids and names, and the GUID rule (lemma
plus part of speech) are unchanged, as they have been since 1.0.0. -
The deck deliberately ships no field for your own notes, which answers
the re-import problem 1.0.0 left open. Anki cannot merge fields on import:
"Always" overwrites every field, and "If newer" overwrites every field of
each note it updates. So any field this package ships can be erased by a
later release, and a field you add cannot be — because once the note type
differs, Anki stops updating those notes at all. A place to write that we
would overwrite is worse than no place.To keep your own notes: add a field to the note type yourself (Tools →
Manage Note Types → Fields). Understand the trade — from then on Anki will
not update these notes, so a later release will add new words but leave your
existing ones as they are. -
Each release now has its own package timestamp. 1.0.0's was frozen, and
because genanki writes it as every note's modification time, Anki's "If
newer" could never see a newer copy: no correction could reach anyone,
whether they had edited a note or not. 1.1.0 is stamped 2026-10-23 and every
release will carry its own, so "If newer" works as it reads. The build is
still byte-reproducible, because the value is pinned inconfig.yamlrather
than taken from the clock. -
New tags. Every tag 1.0.0 set is still set. Added: one
flag::tag per
gloss flag —flag::vulgar(89 notes),flag::bp-leaning(115),
flag::name-like(148),flag::uncertain(318),flag::archaic(13) — so
a learner can find or suspend them; andgloss::reviewedon the 500 notes a
native speaker has passed a verdict on. -
Importing 1.1 over 1.0 will not work cleanly: delete the twenty 1.0 decks
first, or import into a fresh Anki profile. The note type has a new field
(Gloss_Source) and 70 notes changed their part of speech earlier in this
release, which changes their identity. A plain re-import leaves the old
notes behind beside the new ones.Deleting the 1.0 decks does lose whatever review history you have built on
them — the scheduling, the intervals, the lapses. There is no way to carry
it across, because the note type has changed. 1.0 was never announced, so
this should affect few people if any; if you are one of them, the choice is
between keeping your history on a deck with no English side, and starting
again on one that has glosses throughout. -
ep_spoken_anki_enrichable.tsvand its 5001–10000 twin now ship filled
in, with the glosses, the examples, theGloss_Sourcecolumn and the new
tags.
The minimal files are unchanged, for anyone who would rather write their
own glosses from the bare list. -
A native speaker has reviewed the top 500 glosses, over two rounds
(eval/gloss_review.tsv). 500 of 500 reviewed: 479 good, 20 fix, 1 wrong,
none junk. Round two's answers are kept exactly as she sent them in
eval/gloss_review_round2_raw.txt; round one's raw text was not kept, which
is why round two's was committed before anything parsed it.Round one covered 257 rows, of which seven had been judged against gloss text
that changed afterwards and so did not count. Round two covered the 243 rows a
paging bug had hidden and the ten whose text had changed. Its paste runs to 498
lines, but only those 253 are new judgements: the other 245 are round-one
answers her browser still held and copied out again. Every line parsed, and
none fell in the optional band below rank 500. Two rows that carry a round-one
verdict produced no line at all, because their part of speech — and with it
their row id — changed between the rounds; both already counted as done.The 200 sampled rows below rank 500 are still unreviewed and the README says
so.Her corrections are published as overrides: 28 rows in four kinds, named
in the newsourcecolumn ofout/glosses.tsv.override-chosen(11) is a
corpus line she picked from the ones she was shown.override-edited(10) is
a sentence she corrected, so no longer verbatim — trás, for instance, where
she put back an accent the subtitle had dropped. Her eleventh was dever,
which had no example until she wrote one; it is held back rather than
published, because the entry is now a verb and her sentence is the noun.
override-english(4) leaves the corpus line untouched and
takes only her correction to the translation.override(3) is a gloss
written by hand with no sentence: caseiro, which the model refuses to gloss,
largo, and de, whose senses she extended.The English on those rows comes from four places: 5 are hers, given with the
sentence; 12 were written here and she confirmed them; 5 she corrected, one
through the confirm list (um) and four through her notes (este, cidade,
tornar, rápido); and 3 were written here to match sentences she supplied
in round two and she has not seen them yet. Each row says which in itsnote.
gloss.source_labelsinconfig.yamlturns eachsourceinto the wording the
deck'sGloss_Sourcefield will show; that field arrives with the glossed
deck, the column is there now. -
Regenerating the review file no longer destroys the review.
write_tsvoverwroteeval/gloss_review.tsvwith the columns it knows
about, which wiped all 257 verdicts the first time the file was rebuilt
after the advisor filled them in (recovered from git). It now merges,
carrying over any column a reviewer has added, matched on lemma and POS,
and warns when a reviewed row has left the sample. Every other generated
file ineval/already worked this way. -
lorqueis dropped: Iorque with its capital I read as a lowercase l,
7,119 occurrences at rank 3,885, invisible to the proper-noun filter
because it is never capitalized. It andlolaus(Iolaus, 731) are in
eval/proper_noun_drops_review.tsv, which gains anotecolumn. A scan
for siblings — an l where a capital I belongs, never capitalized, whose
I-form is a frequent dropped name — finds 121 in the corpus, but every one
apart from these two occurs 74 times or fewer, far below the list. The two
worth knowing about that the scan cannot see arelgreja(595) and
lnglês(141), whose I-forms are ordinary words rather than dropped names. -
The proper-noun filter measures capitalisation per reading
(fixes.split_aware_proper_nouns). The plain ratio mixes every reading of a
surface together, so a name hiding inside a common word can never reach the
threshold. Where the per-occurrence splitter has divided a surface, the
name reading is now measured on its own.This needs three guards, and without them it is worse than useless. The
PROPN reading must have at least 3 votes and 10% of the surface's votes, and
the capitals must fit inside it. Without them a single stray PROPN tag
collapses the denominator and the ratio saturates at 1.00: the first run
droppedfora,tratar,estado,soldado,dado,espada,privado,
sagrado,gelado,temporada,embaixada,condadoanddeputadoas
names. With the guards, 36 lemmas are re-measured and five cross the
threshold:antártida,prelado,prado,bushidoandunidas(as in
Nações Unidas). Three entries leave the list —lorque,unidas(3,330)
andprado(2,457) — three enter at the bottom, and no count changes at
all.The
sãoandnovaresidues survive, and that is the right answer.
Measured on their PROPN reading alonesãoreads 0.21 andnova0.76,
against a threshold of 0.85 and the 1.00 that names like john and maria
sit at. What is left of them after the tag split is mostly the common word
the tagger mislabelled, not the place — so the remedy fornovais a better
vote share, not the filter. -
largohas a gloss override: "wide, broad; (noun) square, plaza", since
the noun sense lost its own row whenlarga's adjectival occurrences pushed
the noun share below the POS-split threshold. -
sãowas published as a noun at rank 86, with 1,117,272 occurrences.
It is the third person plural of ser — "they are" — and Stanza says so in
45 of its 50 sampled sentences. It was capitalized in only 2% of its
mid-line occurrences, so it is hardly ever the São of São Paulo.The cause was a clause in the consistency rule.
pos.consistentrejects a
VERB or AUX tag whose lemma is a function word, which is right for the POS
column — a preposition tagged VERB is noise — but ser, ter, ir and
este are function words and inflected verbs at the same time. So all 45
AUX votes were thrown away, and the 5 that read the word as a name decided
the lemma.é,somosanderamescaped only because every one of their
votes was dropped, which triggers the "if all are dropped, all count"
fallback.sãohad five survivors, so it did not.Deciding which lemma an occurrence belongs to now uses
pos.consistent_shape, the same rule without that clause, andsãoand
novajoinfixes.ambiguous_formsso their counts are divided by tag:entry before after ser (verb) 21,334,148 22,339,499 são(noun)1,117,272 111,921 novo (adj) 387,344 491,234 nova(noun)216,438 112,548 ser was undercounted by 4.7%. Four counts changed and nothing else: no
entry entered or left the list, the token total, the gold score (98.5%),
the quality gate (0 suspects) and the ser/ir fingerprint are all unchanged.The residues are still not clean entries.
sãokeeps 111,921
occurrences at rank 558, glossed "Saint (before male names); they are" — a
mixture of the name and what the tag sample did not catch.novakeeps
112,548 at rank 553 and is glossed as the feminine adjective, which means
those occurrences probably belong to novo as well: the vote was 26–24, and
the proper-noun side is inflated by Nova Iorque, which already has an
entry of its own at rank 1,050. Neither residue is filtered as a proper
noun, because the capitalization ratio is measured over every occurrence of
the surface, verb uses included. -
largaandlargasare divided between largar and largo by tag
(fixes.ambiguous_forms). The surface is both the verb (larga isso!) and
the feminine adjective (uma rua larga), and the whole 55,702 occurrences
were going to largar — which, once the enclitic fix pushed the count up,
published alargar/adjrow at rank 1,063. The per-occurrence splitter
now divides the count the way it divides fomos between ir and ser:
largar 111,474 and largo 23,636, conserving the total.This needed one more thing to be right.
resolve_jointtook Stanza's lemma
at face value for a listed ambiguous form, and 21 of 50 sampledlarga
sentences come back with lemma largo tagged VERB — Stanza reading "eu
largo" as a form of the adjective. Counting those would have moved most of
a verb's occurrences onto largo, so the consistency rule that already
filters the lemma votes now filters these too. It applies to the ambiguous
branch only: the participle branch derives its lemma from the surface
rather than from Stanza, and filtering it moved 63 entries and 324 counts.Effect: two entries out (
largaradj,largonoun), two in at the
bottom (obediente,hebraico), two counts changed, and nothing else.
Token total, gold score (98.5%), quality gate (0 suspects) and the ser/ir
fingerprint are all unchanged. One consequence worth knowing:largoused
to be published twice, as a noun (o largo, a square — 4,986) and an
adjective (2,276). With 16,374 adjectival occurrences added, the noun share
falls to 21%, below the 25% POS-split threshold, solargois now a single
adjective row and the "square" sense no longer has one of its own.
Changed
-
Every contraction now takes the part of speech of what it contracts
(fixes.contraction_pos,pos.contraction_pos). Stanza expands do,
nas, pelo and the rest into two words before tagging, so the surface
token collected almost no usable evidence: 28 contraction entries read
unk, and those with a stray tag read worse —peloas a noun,desteas
a verb,aoas a conjunction,contigosplit across three parts of
speech. Each form now gets one decided answer, by what it is made of
rather than by the paradigm it sits in:group POS forms preposition + article or demonstrative det58: do, nas, pelo, num, deste, naquela, noutro, … preposition + pronoun pron19: dele, dela, deles, delas, nele, nela, neles, nelas, disto, disso, daquilo, nisto, nisso, naquilo, àquilo, comigo, contigo, connosco, convosco preposition + adverb of place adv3: daqui, daí, dali doanddaalready readdeton the tagger's own evidence; the rest of
the determiner group now agrees with them. The pronoun and adverb groups
are the ones the tagger would never have got right and a single blanket
answer would have got wrong.70 rows changed part of speech and
unkrows fell from 106 to 78.
Because the threecontigorows and four other split entries merge into
one, three entries move into the list from just below rank 10,000
(carnívoro,decretar,lótus). Word lists, counts and ranks are
otherwise unchanged; the gold-set score (98.5%) and the quality gate (0
suspects) are unaffected.This changes Anki note identity for those 70 notes. A note's GUID
comes from its word and part of speech, so re-importing over 1.0.0 adds a
second note for each contraction whose POS changed and leaves the old one
behind. Deleting the twenty decks before importing is the clean path for
anyone who has not yet started reviewing.