Request access to the 4FACTORS Arabic samples

These samples are free for research and non-commercial evaluation. Tell us briefly what you are building — it helps us understand what the field actually needs.

Log in or Sign Up to review the conditions and access this dataset content.

4FACTORS — Palestinian Levantine Conversational Sample

50 native-written question–answer pairs in spoken Palestinian Levantine Arabic, each with an English gloss. This is a public demonstration sample from 4FACTORS, a producer of native, human-verified Arabic training data.

What this is

Real conversational exchanges — the kind of thing people actually say in shops, clinics, taxis, and at home — written from scratch by a first-language Palestinian speaker. Every line was reviewed against a written checklist before release.

  • Not scraped from the web.
  • Not machine-translated from English.
  • Not synthetically generated by a language model.

For a producer of training data in 2026, provenance is the whole point: synthetic Arabic is free and abundant, and models trained on model output degrade. What has value is data a human actually wrote — and can be shown to have written.

Fields

Field Description
id Stable identifier (pal-lev-NNN)
question_ar The question turn, in Palestinian Levantine Arabic
answer_ar The answer turn, in Palestinian Levantine Arabic
translation_en Plain English gloss of both turns, separated by /
domain Everyday setting (shopping, health, transport, government, social, …)
variety Arabic variety — Palestinian Levantine throughout

Example

question_ar answer_ar translation_en domain
السلام عليكم، بغلّبك، في عندك خبز وجبنة بيضاء؟ اه في، كم بدك خبز وجبنة؟ Hello, sorry to bother you — do you have bread and white cheese? / Yes, we do. How much do you want? Shopping
دكتور، في خافض حرارة للولد؟ اه، في. بس قديش عمر الولد؟ Doctor, do you have a fever reducer for my child? / Yes, we do. But how old is the child? Health

How it was made

  1. Scoped brief in the writer's own dialect, with explicit prohibitions on copying, translating, and AI generation.
  2. Native composition — the speaker writes from lived situations, not from prompts scraped off the web.
  3. Review against a checklist — no real names, no real businesses, no copyrighted text, honest domain labels, natural spoken register.
  4. Normalisation and English glosses added by the 4FACTORS team.

The full 150-item public sample (Palestinian Levantine, Egyptian, and Modern Standard Arabic), including the reviewer's correction log for the Egyptian set, is described at the project page below. All three varieties are published on Hugging Face and Kaggle under CC BY-NC 4.0.

Scope and limitations

  • 50 items — a demonstration sample, not a training-scale corpus.
  • Single writer, single variety. Not balanced for speaker demographics.
  • English glosses are plain functional translations for evaluation, not literary translation.

Commissioned datasets

4FACTORS collects Arabic data to a client's schema, in the varieties and volumes they need — Levantine, Egyptian, Gulf, Maghrebi, MSA, and others on request. Commissioned datasets are delivered with full IP transfer and English glosses included.

Licence

Released under CC BY-NC 4.0 — free to evaluate and use non-commercially, with attribution. Commercial training use is available via a commissioned or licensed dataset; contact us. All contributors are compensated and have consented to publication. No personal data. No real names, brands, or copyrighted source text.

Archived deposit

This sample is part of an archived corpus (all three varieties, 150 items) published on Zenodo together with the full provenance record — contributor brief, review process and SHA-256 checksums:

DOI: 10.5281/zenodo.21380048

Citation

@dataset{4factors_arabic_2026,
  author    = {Poljanc, Andrej and {4FACTORS GmbH}},
  title     = {4FACTORS Arabic Native-Written Sample Corpus: Modern Standard
               Arabic, Palestinian Levantine and Egyptian (150 items)},
  year      = {2026},
  version   = {1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21380048},
  url       = {https://doi.org/10.5281/zenodo.21380048},
  note      = {Native-written, human-verified. CC BY-NC 4.0.}
}
Downloads last month
55