Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
78 lines (56 loc) · 3.18 KB

CONTRIBUTING.md

File metadata and controls

78 lines (56 loc) · 3.18 KB
Raw

Contributing

Thank you — this list is only useful if it stays accurate.

There are three ways to help, in descending order of usefulness.

1. Verify an entry (most valuable)

Entries marked ¹ in the README have verified: false: their numbers come from a secondary source (a survey paper, another list) rather than the primary release. They may be wrong.

To verify one:

  1. Open the dataset's primary source — the official page, the paper, or the Hugging Face card from the authors (not a mirror).
  2. Confirm hours, sample_rate, labels, license, and access.
  3. Fix anything wrong in datasets.yaml, set verified: true, and open a PR linking the exact source you checked.

If the primary source genuinely does not state a figure, leave it null rather than guessing. null is honest; a made-up number is worse than nothing.

2. Add a dataset

Open an issue or send a PR adding an entry to datasets.yaml.

Inclusion criteria — a dataset belongs here if it:

  • contains Vietnamese speech audio (including Vietnamese minority/ethnic languages and Vietnamese-accented speech), and
  • is publicly obtainable — open download, free registration, or a documented request process. Datasets available only through a commercial contract do not qualify.

Purely text corpora, and models without an accompanying dataset, belong in awesome-vietnamese-nlp instead.

Write the notes field like you are warning a colleague. This is the part of the repo that cannot be found elsewhere. Good notes say what the data actually sounds like, where the labels break down, and what will bite you at 3am:

✅ "Utterances average ~3s and there is no punctuation or capitalisation; an independent analysis found a substantial fraction fails quality filters."

❌ "A high-quality large-scale dataset for Vietnamese ASR research."

3. Report a problem

Broken link, licence changed, dataset taken down, numbers that do not match the source — open an issue. Dead links are a real problem for lists like this and reporting one takes 30 seconds.

Ground rules

  • datasets.yaml is the source of truth. Never edit the generated tables in README.md by hand — your change will be overwritten.

  • After editing, run both scripts locally:

    pip install pyyaml
python scripts/validate.py      # schema check
python scripts/gen_readme.py    # regenerate README tables

    CI runs the same two checks on every PR and will fail if the README is stale.

  • No promotional language. Not for datasets, not for the labs that produced them. If a corpus has known problems, they go in notes — even (especially) if it is one of ours.

  • One dataset per PR keeps review easy.

A note on scope and ethics

Several large Vietnamese corpora are scraped from YouTube and pseudo-labelled. This list catalogues them because pretending they do not exist helps nobody, but it also records what they are, so that people can make informed decisions. If you know of a rights or consent issue with a listed dataset, please open an issue — that information belongs in notes too.