Repository navigation
Releases
1.1.0 — Spoken European Portuguese frequency list, with English glosses
⚠️ 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 attest...
1.0.0 — Spoken European Portuguese frequency list
The first public release: 10,000 words of European Portuguese film and TV
dialogue, ranked, with a ready-to-import Anki package. Lemmas come from a
neural tagger checked against a dictionary and score 98.5% on a
hand-reviewed sample; the part of speech is the tagger's, and ser and
ir are separated by a rule measured against a native-speaker-labeled
sample.
Overlap with the April list is 79.0% of band 1 (ρ 0.944) and 63.3% of
band 2 (ρ 0.851). Against the v0.9 checkpoint the word lists barely change;
what changed there was the POS column and, through POS splitting, which
rows fill the last places of each band.
Known issues.
-
No glosses or example sentences. The
Gloss_EN,Example_PTand
Example_ENfields ship empty, for you to fill in; until a note has a
gloss it has no English → Portuguese card. Glosses are planned for 1.1. -
Six accent pairs are accepted as known typos and appear as two
entries each:camera/câmera,frigorifico/frigorífico,
amen/ámen,bla/blá,mafia/máfia,karate/karaté. They sit
below the 10× folding ratio, and the quality gate accepts them by name. -
The
ser/irrule is right on 95.9% of the verb uses it decides on
the development sample and 93.2% (68 of 73) on the held-out sample,
abstaining on 8–9%. Its errors all call aniruseser, soiris
still slightly undercounted. One adverbial fora in twenty reaches the
rule at all (lá fora, mistagged by Stanza), affecting about 1% of
foratokens; left for a later round rather than tuned away against the
test set. -
Re-importing a later
.apkgmatches notes by word and part of
speech, and Anki's "Update notes" option then replaces every field,
glosses and examples a learner has typed in included; choosing "Never"
keeps them but freezes ranks. Deferred to 1.1, when glosses ship. -
A ready-to-import Anki package,
out/EP_Spoken_Frequency.apkg: one
note type, twenty decks of 500 words numbered in rank order
(01 · 1–500…20 · 9501–10000), a Portuguese → English card
for every note, and an English → Portuguese card that appears only once
a gloss is added. Notes are identified by word and part of speech, so a
later release imported over this one updates notes instead of
duplicating them. The TSV files remain for building your own note type. -
Part of speech from Stanza. The POS column -- renamed from
pos_guesstopos, and the Anki field fromPOS_guesstoPOS-- is
now the tagger's majority vote over each word's sample sentences,
replacing the rule heuristic reconstructed from the April list.
Prepositions are never counted as conjunctions (Universal Dependencies
tags para fazer as one); the list is in config.yaml. A word the tagger uses
as two parts of speech gets one row for each, with its count divided the
same way participle counts are divided across lemmas: 311 words now have
more than one row (aas article and preposition,que,este,
morto), so the 10,000 rows hold 9,390 distinct words. 148 entries
without usable tags, mostly contractions, keep the heuristic. -
Larger samples for frequent words. Word forms with 10,000+
occurrences are now read in 50 sentences instead of 5, so their splits
move in 2% steps instead of 20%. -
One tagging pass. Lemmas, participle splits and POS now come from a
single Stanza pass over about 500,000 sentences, and a lemma vote is
discarded when Stanza's lemma and tag contradict each other (saia
given the lemma saia but tagged VERB). Gold-set accuracy is unchanged
at 98.5%.váadded to the override table (Stanza read it as ver). -
serandirare now split by a rule. v0.9 said the tagger split
them by context; it does not. The larger sample shows it assignsfoi,
fui,fomosandforamto ser almost regardless of context, all 50
sampled uses offuiandfomosincluded (fui ao Paquistão). Stanza
now only decides whether an occurrence is a verb, which keeps the adverb
fora ("outside") out, and a next-word rule (scripts/serir.py) picks
ser or ir. Scored against 100 random corpus lines checked by hand
(eval/ser_ir_sample.tsv): 95.9% right on the verb uses it decides,
abstaining on 9%; none of the 19 adverbial fora lines reaches it. On a
held-out sample of another 100 lines labeled afterwards by a
native-speaker Portuguese teacher (eval/ser_ir_sample2.tsv), 93.2%
(68/73), abstaining on 8% —
the honest estimate, since the rule was never tuned against it.
Abstentions, and non-verb tags on forms that are always verbs (Stanza
tags 6% offoias a conjunction in clefts like Foi por isso que),
take the form's own ser/ir ratio. The build applies the rule only
while a passing score exists for its current version. Per form, ir is
now 26% offui, 29% offomos, 11% offoramand 10% offoi. Ser
goes from 21.5M tokens to 21.3M and ir from 8.9M to 9.0M:fuiand
fomosmove a lot, butfoi, the biggest form by far, is mostly the
copula.
v0.9 — 2026-09, rebuilt pipeline, human-checked (private checkpoint)
The first version whose pipeline is in this repository. python -m scripts.build regenerates out/ from the corpus, byte for byte, with the
pinned library versions.
Overlap with April. 79.5% of the April top-5,000 is still in the top
5,000 (Spearman ρ 0.948 on the shared words); 62.4% of the 5,001–10,000
band (ρ 0.879). Most of the movement is deliberate — the April list counted
the same word several times, and this release does not.
Why ranks moved
- One word, one entry. April's lemmatizer left inflected forms and
spelling variants as separate entries:naoat rank 1004 besidenão,
achabesideachar,chegabesidechegar,exactamenteand
exatamentecounted apart. They are now merged, which raises the
surviving headword and frees the rank the duplicate occupied. - A better lemmatizer. simplemma alone was replaced by Stanza, reading
each word in real sentences, with a dictionary check and simplemma as the
fallback: 98.5% correct on a human-reviewed sample, against 90.4% for
simplemma alone. serandirseparated — later found not to work; see
Unreleased. April assignedfoi,fomos,foramwholesale toir.
They were split by the tagger:irfell from 11.2M tokens to 9.5M, and
serrose from 20.6M to 22.9M.- Enclitics split. April counted
dá-meandvou-meas single words
(its README said otherwise). They are now split into verb + pronoun, so
vai-te emborano longer creates its own entries, and object pronouns
(merank 15,te23,lhe70) are counted in full. - Lemmatization conventions, written down in
eval/conventions.md: contractions are entries,
nouns for people keep the feminine (senhora,rapariga), diminutives
and comparatives stay separate, and headwords use the post-1990
spelling. ais now separate fromo. April folded the preposition/article
aintoo;odrops from 58.1M tokens to 29.9M, andarises from
rank 8281 to rank 3.- Proper nouns removed properly. April kept
john(rank 449),mr,
charlie,peterandjoeas vocabulary. Names are now removed by capitalization, with
every dictionary word removed at 5,000+ occurrences reviewed by hand —
deus,srandnatalstay. - Corrections to April's exclusions.
cara("face") was excluded as
Brazilian by mistake and is back (rank 340); the unaccentedvoce
slipped past the BP filter and is now caught. - Junk removed. The subtitle watermark
pt-subs rips, the MWEd c,
English plurals (zombies) and one- or two-letter fragments are gone.
Also new
- A 400-word gold set, human-reviewed (eval/lemma_gold.tsv),
and a quality gate that fails the build on suspected duplicate entries.
The release passes it. Six missing-accent pairs below the folding
threshold are accepted as known typo pairs and listed in the README. - Reports in reports/: comparison with April, the quality gate
with every accent fold, and the lemmatizer comparison. - Multi-word expressions: 302 in the top 10,000 (April: 290).
- Token count 636,214,452 (April: 623,920,347), because split enclitic
pronouns now count as tokens of their own.
Unchanged
The pos_guess column was still a rule-based heuristic, reconstructed
from the April list.