Aceh NLP

A language lives where it is spoken.

Acehnese is carried in kitchens, on fishing boats and over coffee, far from the data that trained today’s models, which learned about Aceh from other people’s choices and inherited their blind spots. We are a small group from different backgrounds, gathered on Pulau Weh by curiosity and care for the language, helping speakers collect their own words and decide what counts as data, who checks it, and what it is for. This is how we begin to decolonise the technology: a language model held by a community becomes a living archive, one you can question, correct and teach, so knowledge keeps moving, as it always has, from one voice to the next.

Evaluation Talk with the model

How it works

View

Think of a family kitchen. The cook learns from one recipe book, and the family decides what goes back into it.

The same loop in standard machine-learning terms.

The recipe book

Dataset

Dataset

  • Elders’ recipesLayer 1: specialistSpecialist-checked. Smallest, most trusted.
  • Neighbours’ recipesLayer 2: everydayEveryday speakers and internet text.
  • Machine-written recipesLayer 3: syntheticSynthetic AI drafts. Largest, least trusted.

A recipe moves up only when a person has tasted it.

A sentence moves up a layer only after human review.

The cook learns the book

Training

Training

The cook

Model

Model

A sealed family recipe

Locked test set

Locked test set

The cook never reads it. It shows whether each new cook is really better.

Held back from training. Scores each new model.

Cooks new dishes

Generate

Generate

The family tastes

Human review: annotation and evaluation

Human review: annotation and evaluation

Elders (specialists)SpecialistsScore 1–5, correct and annotate.

Everyone elseEveryday speakersFor now, score 1–5, correct and annotate too. Later: compare two answers to one question, and give the better one +1 and the other −1.

What passes goes into the book.Reviewed data returns to the dataset.

People behind it

Six of us, living and experimenting together on Pulau Weh — which is how the project began. We come from different places and backgrounds:

USMalaysiaChinaAcehIndonesia

Join the project

It takes many kinds of help, not only reviewers. All roles are volunteer for now.

  • Teach or study Acehnese

    Send us links to Acehnese lessons, such as YouTube videos or teaching material.

  • Speak Acehnese

    Score and correct AI-drafted sentences.

    Start evaluating

  • Live on Pulau Weh

    Help us work respectfully with communities.

  • Record and transcribe

    Lend equipment or train as an interviewer. Recording starts once consent is ready.

  • Share writing or translate

    Acehnese text you have the right to share, or translation between Acehnese and Indonesian.

  • Build software

    Review tools, data and training.

  • Offer compute or funding

    GPU time, cloud credits and grants, for when training is ready.

  • University or legal

    Permits, consent and data protection.

  • Design, write, film

    Help people understand the project honestly.

To take part, write to hello@aceh-nlp.net