atlasia/lid-dataset-base

LID Dataset Base

This is the base dataset used for Moroccan Darija language identification. The job it's built for is: given a piece of text, decide whether it's Moroccan Darija or something else (darija vs other). On top of that binary label, every row also carries a finer-grained dialect tag (46 of them) so the same data can be sliced for dialect-level analysis, not just the binary task.

It's a merge of many datasets mentionned in source column, deduplicated, cleaned the same way on both sides, re-labeled with a single consistent dialect scheme, and re-split.

~2.05M rows, 46 dialects/languages, 80/10/10 train/val/test.

At a glance

Total rows	2,048,654
Train / Validation / Test	1,638,923 / 204,865 / 204,866 (80 / 10 / 10)
Binary label	darija 963,556 (47%) · other 1,085,098 (53%)
Distinct dialects	46 (all present in every split)
Dialect families	11
Source corpora	67
Text in Latin script	896,097 rows (44%)
Median cleaned length	12 words (mean 38, max 710)

Columns

Column	What it is
src	The original source tag. This is what the dialect was derived from.
text	The raw sentence, untouched.
text_clean	The cleaned text. Same cleaning service applied to both source datasets so they're comparable. This is also the dedup key.
label	The binary target: darija or other.
dialect	Fine-grained dialect / language, e.g. moroccan_darija, algerian, egyptian, french.
dialect_family	Coarse grouping, e.g. Maghrebi, Levantine, Non-Arabic.
country	Country the dialect maps to, where it makes sense (null for non-Arabic languages and MSA/English).
has_latin_script	"true" / "false" — whether the text contains any Latin characters. This is how arabizi shows up, since the dialect column itself no longer separates arabizi from Arabic script.
word_count_clean	Word count of text_clean.

One thing to know about the label

label == "darija" means Moroccan Darija specifically. It lines up exactly with dialect == "moroccan_darija" (both are 963,556 rows). Everything else including the other Maghrebi dialects like Algerian, Tunisian, Libyan is other. So if you train on the binary label, you're training a Moroccan Darija detector, not a "MAR Darija vs. the world" detector. Use the dialect column if you want anything more granular than that.

How the dialect was decided

The dialect is not guessed from the text. It's derived from the source tag, which encodes where each sentence came from.

• madar:msa → msa, madar:cairo → egyptian (MADAR corpus, by city)
• arabic_tweets:EG → egyptian, arabic_tweets:LB → lebanese (ISO country code)
• flores:arz → egyptian, flores:ary → moroccan_darija (ISO-639-3 language code)
• atlaset → moroccan_darija, elner_dz → algerian (named single-dialect corpora)

After deriving the raw dialects, a normalization step is applied to them so the labels are consistent:

• Collapsed the fine-grained variants: e.g. palestine, palestinian_arabic_surif_falahi_dialect, palestinian_nabulsi_urban_dialect all became palestinian; lebanon / lebanese_standard_dialect became lebanese.
• Dropped the script split from the dialect column arabic_moroccan and arabizi_moroccan are both just moroccan_darija now. The arabizi/Arabic information lives in has_latin_script instead, so nothing is lost.

What's in it

Dialects

moroccan_darija is ~47% on its own, and with Algerian the Maghrebi family is most of the data. The long tail is a wide spread of other Arabic dialects plus non-Arabic languages (these are the negatives that make the other class realistic).

Full dialect table (46)

dialect	family	country	label	rows	share
moroccan_darija	Maghrebi	Morocco	darija	963,556	47.03%
algerian	Maghrebi	Algeria	other	418,492	20.43%
english	Universal	Standard	other	100,156	4.89%
egyptian	Egyptian	Egypt	other	49,518	2.42%
lebanese	Levantine	Lebanon	other	44,534	2.17%
palestinian	Levantine	Palestine	other	28,127	1.37%
levantine	Levantine	Jordan	other	27,271	1.33%
libyan	Maghrebi	Libya	other	26,231	1.28%
msa	MSA	Standard	other	20,520	1.0%
german	Non-Arabic	—	other	18,262	0.89%
portuguese	Non-Arabic	—	other	18,238	0.89%
turkish	Non-Arabic	—	other	18,223	0.89%
dutch	Non-Arabic	—	other	18,178	0.89%
italian	Non-Arabic	—	other	18,164	0.89%
spanish	Non-Arabic	—	other	18,101	0.88%
french	Non-Arabic	—	other	18,000	0.88%
danish	Non-Arabic	—	other	16,957	0.83%
vietnamese	Non-Arabic	—	other	16,934	0.83%
czech	Non-Arabic	—	other	16,921	0.83%
swedish	Non-Arabic	—	other	16,912	0.83%
hungarian	Non-Arabic	—	other	16,896	0.82%
polish	Non-Arabic	—	other	16,896	0.82%
indonesian	Non-Arabic	—	other	16,863	0.82%
tunisian	Maghrebi	Tunisia	other	16,374	0.8%
sudanese	Sudanese	Sudan	other	15,413	0.75%
syrian	Levantine	Syria	other	14,661	0.72%
jordanian	Levantine	Jordan	other	12,876	0.63%
gulf	Gulf	Qatar	other	10,945	0.53%
saudi	Gulf	Saudi Arabia	other	9,550	0.47%
omani	Gulf	Oman	other	7,364	0.36%
mauritanian	Maghrebi	Mauritania	other	6,624	0.32%
yemeni	Yemeni	Yemen	other	6,144	0.3%
banjar	Non-Arabic	—	other	2,009	0.1%
acehnese	Non-Arabic	—	other	2,008	0.1%
minangkabau	Non-Arabic	—	other	2,007	0.1%
urdu	Non-Arabic	—	other	2,003	0.1%
dari	Non-Arabic	—	other	1,991	0.1%
kashmiri	Non-Arabic	—	other	1,991	0.1%
iraqi	Mesopotamian	Iraq	other	1,990	0.1%
persian	Non-Arabic	—	other	1,975	0.1%
kanuri	Non-Arabic	—	other	1,960	0.1%
pashto	Non-Arabic	—	other	1,901	0.09%
azerbaijani	Non-Arabic	—	other	1,830	0.09%
romanian	Non-Arabic	—	other	1,252	0.06%
catalan	Non-Arabic	—	other	1,242	0.06%
najdi	Gulf	Saudi Arabia	other	594	0.03%

Families

Label balance & script

The binary classes are close to even (47/53). Just under half the rows are in Latin script, that's mostly arabizi (Darija written with Latin letters and numbers) plus the non-Arabic languages.

Text length

Short text is the norm, half the rows are 12 words or fewer. There's a long tail up to ~700 words.

Loading it

from datasets import load_dataset

ds = load_dataset("atlasia/lid-dataset-base")
print(ds)
# train / validation / test, with the columns above

ds["train"][0]

The splits were made with stratification on dialect, so the dialect and label proportions match across train/val/test, and all 46 dialects appear in every split. Dedup was done on text_clean before splitting, so there's no text_clean leakage between train, validation, and test.

Things to keep in mind

• The labels are source-derived, not human-annotated per sentence for some sources. A tweet from arabic_tweets:EG is labeled Egyptian because that's where it came from. That's accurate at the corpus level but a small piece of MSA or off-dialect sentence inside an otherwise-Egyptian source still gets the Egyptian tag. Don't treat the fine-grained dialect as gold per-row truth.
• darija = Moroccan only (see above). Algerian and the other Maghrebi dialects are negatives.
• It's imbalanced. Moroccan + Algerian dominate; the rarest dialects (najdi, romanian, catalan) have only a few hundred rows. Sample or weight accordingly.
• A couple of country / family values are a bit rough, e.g. the coarse levantine bucket lands under Jordan, and english is split between the Non-Arabic and Universal families depending on which source it came from. The dialect and label columns are the reliable ones.