Portuguese-NLP: a curated index of Portuguese datasets, not a toolkit
List of resources and tools developed with focus on Portuguese.
At a glance
- What is it?
- ajdavidl/Portuguese-NLP is a README-only list of datasets and resources for Portuguese-language NLP. It solves discovery, not processing, and its value depends entirely on how much you trust someone else's link list.
- Who is it for?
- Use Portuguese-NLP if you are scoping a Portuguese-language project and need to find candidate corpora before committing to a collection strategy; the README's links to ASSIN, CORAA, Carolina, BrWaC and the Hugging Face mirrors are a faster starting point than searching Hugging Face blind. Do not use it if you need a library, a tokenizer, a model checkpoint or anything you can import, because the repository contains only README.md.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 108 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on October 11, 2026, and from our analysis. They are not legal advice.
Editorial analysis
What the Portuguese-NLP list actually is
The repository holds one file, README.md. There is no package manifest, no source directory, no test suite and no build configuration. That is not a defect in the project; it is the project. Portuguese-NLP is a link index, in the same genre as the various awesome-* lists, and the README describes itself as a "List of resources and tools developed with focus on Portuguese."
The audience is narrow and specific. It is for an engineer or researcher who has decided to work on Portuguese text or speech and now has to answer a prior question: what data already exists, and where does it live? That question is genuinely expensive in Portuguese, because the resources are scattered across university pages, Linguateca, Kaggle, Zenodo, Hugging Face and individual GitHub repositories. A single page that collects them has real value even though it computes nothing.
What it is not is a toolkit. If you arrived expecting a tokenizer, a lemmatizer, a spaCy pipeline or a pretrained checkpoint, you will not find one here. The README's own framing is resources and tools, but the entries are overwhelmingly datasets, with a handful of links to project repositories that may themselves contain models. The list points; it does not provide.
How the index is organised and what the entries contain
The README is a flat bulleted list under a Datasets heading. There is no table of contents, no categorisation by task, no grouping by language variant and no indication of size or licence for most entries. Each line follows a loose pattern: a bracketed name, a link, and a short gloss. Some glosses are a single phrase, some are a sentence, and a few include a second link to a Hugging Face mirror.
The glosses carry the only selection signal available. For example, the entry for ASSIN is described as "a dataset with semantic similarity score and entailment annotations" and points to both a NILC page and a Hugging Face dataset. The entry for CORAA is described as a "dataset for Automatic Speech Recognition." The entry for Carolina points to a USP site and a Hugging Face mirror. Where a Hugging Face mirror exists, that is usually the fastest route to the data, because the mirror page carries the loading code and the licence metadata that the list itself omits.
The ordering is alphabetical by display name, which means task-adjacent resources sit far apart. ASSIN and ASSIN 2 are adjacent because of the name, but a sentiment corpus and a speech corpus are not grouped. If you are looking for, say, everything usable for named entity recognition, you will read the whole list and follow the glosses. That is a minor cost for a list of this size, but it is a cost, and it is the main thing a table would have fixed.
Getting a first dataset from the Portuguese-NLP list
There is nothing to install. The repository has no package, so any instruction to run pip or npm against it would be invented. The README does not give install steps for itself, and it does not give loading snippets for the datasets either; those live on the destination pages. The repository does give the clone URL, which is the only command it supports:
git clone https://github.com/ajdavidl/Portuguese-NLP.gitAfter that you have README.md and nothing else. Grep it for the task you care about, for instance speech or sentiment, and follow the links that match. The practical first move is to pick an entry that has a Hugging Face mirror, because the mirror page is where the actual loading instructions are. The README names several such mirrors, among them assin, assin2, carolina-c4ai/corpus-carolina and dominguesm/canarim. Copy the identifier from the mirror page rather than from the list, since the display names and the mirror identifiers are not always identical.
What you should see, once you are on the mirror page, is a dataset card with a licence field and a usage snippet. If the identifier does not resolve, that is a signal that the list entry has drifted, which is the normal failure mode of a link index.
Where the list breaks down as a decision tool
The most serious gap is licensing. The README records no licence for any entry, and the repository itself has no licence file in its top-level entries, which are README.md only. For a corpus you intend to ship inside a product, that is the first thing you need and the last thing this list gives you. Some destinations are permissive, some are research-only, and some are gated behind a form. You will not learn which from the bullet.
Size and composition are also missing. The gloss for the Brazilian Portuguese Literature Corpus states "3.7 million word corpus of Brazilian literature published between 1840-1908," which is unusually informative. Most entries say nothing comparable, so you cannot rank candidates by volume without visiting each one. For a task like pretraining or fine-tuning, volume is often the deciding factor, and the list does not help you sort.
There is also a maintenance question that the format invites. A link index decays: pages move, university hosts retire directories, and a bullet that resolved last year may 404 today. The last push to this repository was on 2026-06-25, so the list has been touched recently, but a push does not guarantee every link was rechecked. Budget time for dead ends. Finally, the list is heavily weighted toward Brazilian Portuguese. European Portuguese resources appear, for instance CINTIL and CETENFolha, but the centre of gravity is Brazil, and a project targeting pt-PT should not treat this as a balanced inventory.
Alternatives: Hugging Face Hub and task-specific corpora
The obvious alternative is to skip the index and search the Hugging Face Hub directly by language tag. The difference in approach is fundamental: the Hub is a live catalogue with licence metadata, dataset cards, download counts and a consistent loading API, whereas Portuguese-NLP is a hand-written page of links. The Hub will surface resources this list has never heard of, and it will tell you the licence before you download.
What the Hub does not do well is cover the older, institution-hosted material. Linguateca corpora such as CETENFolha and CHAVE, the C-ORAL-BRASIL spoken corpus, and the USP Carolina corpus are not Hub-native in the same way, even where a mirror exists. Those are exactly the resources a curated list is good at surfacing, because they are hard to find by tag search and easy to miss.
A second alternative is a task-specific starting point rather than a general index. If you are doing automatic essay scoring, the AES ENEM dataset and Essay-BR entries are the relevant ones; if you are doing clinical NER, ClinicalNER is the entry; if you are doing speech, CORAA. Going straight to the task-specific resource is faster than reading a general list, at the cost of not knowing what else exists. The list's real function is the survey, not the lookup.
Maintenance, licensing and what the list does not tell you
The repository is not archived, and the last push was on 2026-06-25. That is recent enough that the list is plausibly still being extended, but the README gives no changelog, no contribution guide and no statement of how entries are added or verified. There is no release history, so there is no version to pin and no upgrade path to plan. Your maintenance cost is the cost of re-checking links when something breaks, which you will discover at the moment you try to use a dataset rather than in advance.
The repository's own licensing is unstated: no licence file appears among the top-level entries. That matters less than it sounds, because a list of links is not the data, but it does mean you have no explicit grant to redistribute the README. The datasets it points to carry their own terms, and those terms vary by destination. Kaggle entries sit behind Kaggle's account and terms; Hugging Face entries carry a licence field on the dataset card; institutional pages may state research-only use in prose. None of that is summarised here, and the list should not be treated as a rights clearance. Check each destination before you build on it.
Editorial conclusion
Use Portuguese-NLP if you are scoping a Portuguese-language project and need to find candidate corpora before committing to a collection strategy; the README's links to ASSIN, CORAA, Carolina, BrWaC and the Hugging Face mirrors are a faster starting point than searching Hugging Face blind. Do not use it if you need a library, a tokenizer, a model checkpoint or anything you can import, because the repository contains only README.md. Before relying on any single entry, open the linked page and check the licence and the access route yourself; the list does not state licences, and several entries point at Kaggle or institutional pages where the terms live on the destination site, not here.
Frequently asked questions
Which AI is the best for learning Brazilian Portuguese?
The list does not evaluate assistants or language-learning products, so it cannot answer this. Its entries are datasets and corpora for building NLP systems, not tools for a learner.
Why is Brazilian Portuguese so nasally?
This is a phonetics question and the list does not address it. The closest relevant entries are spoken corpora such as C-ORAL-BRASIL and CORAA, which contain speech data rather than explanations.
How do Brazilians say "let's go"?
The list does not cover vocabulary or translation. It indexes datasets, including parallel corpora such as Capes, but it does not provide phrase-level answers.
Why do Brazilians say "tudo bem"?
The list gives no linguistic explanation of expressions. It is a directory of Portuguese-language resources, not a reference on usage.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/ajdavidl-portuguese-nlp)