Dihya-22M

A 22.0M-parameter (21,999,616 published) byte-level Conv-Transformer that upgrades Dihya-5M across the entire Berber dialect continuum in both Latin and Neo-Tifinagh scripts, North African Arabic dialects in Arabic script and Latin Arabizi, Modern Standard Arabic, and major world languages: twenty-two classes across five scripts.

On a stratified held-out set of 14,247 sentences across 36 (language, domain) cells it reaches 91.90% accuracy and 0.9180 macro-F1 over all twenty-two classes. Every single class achieves $\ge 71.4%$ recall, and its worst group recall is 63.33% on the scarcest Shawiya cell.

Against Dihya-5M, it expands the inventory from seven classes to twenty-two, adds Neo-Tifinagh, Algerian Darija, Moroccan Darija, and Arabizi, lifts overall accuracy from 85.52% to 91.90%, and resolves the two central weaknesses of the 5M checkpoint:

1. Kabyle precision rises from 0.626 to 0.900 while keeping recall at 0.925. Dihya-5M traded precision for recall, misrouting 37% of its Kabyle predictions to sibling text; Dihya-22M ends that trade-off.
2. Central Atlas Tamazight recall rises from 0.660 to 0.948, and its confusion into Kabyle drops from 16.0% down to 2.58%. Tarifit confusion drops from 6.7% to 1.38%, and Tamasheq from 7.3% to 0.16%. Across Neo-Tifinagh varieties, confusion to Kabyle is $\le 2.12%$ (0.00% for Tamazight, Tarifit, and Tamasheq).

It ships with a split-conformal quantile ($\hat{q} = 0.9526$ at 99.0% nominal coverage, reaching 98.28% empirical coverage), so it returns a set of labels rather than one when text lies on the continuum and carries no distinguishing marker.

The task

Language identification across the Berber dialect continuum and North Africa confronts three structural challenges:

1. Morphological discrimination over shared roots: Amazigh varieties share substantial core vocabulary; distinguishing evidence is predominantly affixal, phonotactic, and clitic. Short sentences lacking dialect-specific markers are genuinely ambiguous.
2. Upstream corpus contamination: Broad-coverage classifiers (such as NLLB's lid218e and GlotLID) fail on this continuum by collapsing sibling varieties into Kabyle—mislabelling 91.3% of Shawiya, 80.7% of Tarifit, 79.3% of Tashelhit, and 73.3% of Central Atlas Tamazight. Unfiltered web and parallel corpora inherit this silent contamination.
3. Multi-script and dialect boundaries: Written Berber alternates between Latin and Neo-Tifinagh orthographies, while North African Arabic spans Arabic script and Latin-digit Arabizi alongside close dialectal proximity between Algerian (arq) and Moroccan (ary) Darija.

Results

14,247 sentences across 36 (language, domain) cells, held out completely disjoint from train (113,974) and dev (7,145).

Against the public identifiers and Dihya-5M

All systems scored in one pass over the held-out test rows:

system	accuracy	classes it can name	Amazigh coverage	Darija & Arabizi
Dihya-22M	91.90%	22 of 22	13 varieties (Latin & Tifinagh)	arq, ary, Arabizi
Dihya-5M	85.52%	7 of 22	6 varieties (Latin only)	none
GlotLID	41.78%	3 of 22	3 varieties (Latin only)	conflated / no Arabizi
NLLB lid218e	28.33%	2 of 22	2 varieties (Latin only)	conflated / no Arabizi

Public identifiers cannot name Shawiya, Central Atlas Tamazight, Tarifit, Tumzabt, or Arabizi. When asked to classify texts in these varieties, their accuracy is zero by construction.

The quantity that matters downstream is how much sibling text ends up labelled Kabyle, since that is what a filter keyed on kab silently ingests:

true language	script	→ Kabyle, Dihya-22M	→ Kabyle, Dihya-5M	→ Kabyle, GlotLID	→ Kabyle, lid218e
Tashelhit	Latin	1.38%	2.0%	17.3%	79.3%
Tarifit	Latin	1.38%	6.7%	90.0%	80.7%
Tumzabt	Latin	1.88%	—	—	—
Central Atlas Tamazight	Latin	2.58%	16.0%	66.7%	73.3%
Shawiya	Latin	21.43%	19.3%	96.0%	91.3%
Tamasheq	Latin	0.16%	7.3%	0.0%	6.7%
Tashelhit	Tifinagh	0.50%	—	—	—
Central Atlas Tamazight	Tifinagh	0.00%	—	—	—
Tarifit	Tifinagh	0.00%	—	—	—
Standard Moroccan	Tifinagh	2.12%	—	—	—
Tamasheq	Tifinagh	0.00%	—	—	—

Central Atlas Tamazight leakage into Kabyle drops sixfold from Dihya-5M's 16.0% down to 2.58%, and Tamasheq drops from 7.3% to 0.16%. In Tifinagh, cross-leakage into Kabyle is virtually zero.

Per class, all twenty-two

The twenty-two-class report, which is what test-report.json holds:

label	language / dialect	script	support	precision	recall	F1
kab_Latn	Kabyle	Latin	800	0.900	0.925	0.912
kab_Tfng	Kabyle	Neo-Tifinagh	800	0.971	0.879	0.923
shy_Latn	Shawiya	Latin	98	0.864	0.714	0.782
mzb_Latn	Tumzabt / Mozabite	Latin	800	0.986	0.963	0.974
shi_Latn	Tashelhit	Latin	800	0.984	0.975	0.979
shi_Tfng	Tashelhit	Neo-Tifinagh	800	0.716	0.744	0.730
tzm_Latn	Central Atlas Tamazight	Latin	659	0.947	0.948	0.948
tzm_Tfng	Central Atlas Tamazight	Neo-Tifinagh	368	0.986	0.927	0.955
rif_Latn	Tarifit	Latin	800	0.975	0.979	0.977
rif_Tfng	Tarifit	Neo-Tifinagh	69	0.938	0.884	0.910
zgh_Tfng	Standard Moroccan Amazigh	Neo-Tifinagh	800	0.678	0.743	0.709
taq_Latn	Tamasheq	Latin	627	1.000	0.994	0.997
taq_Tfng	Tamasheq (Tawallammat)	Neo-Tifinagh	265	1.000	0.989	0.994
arq_Arab	Algerian Arabic (Darija)	Arabic	788	0.738	0.898	0.811
arq_Arabizi	Algerian Arabic (Arabizi)	Latin/Numerals	223	0.911	0.919	0.915
ary_Arab	Moroccan Arabic (Darija)	Arabic	800	0.846	0.781	0.812
ary_Arabizi	Moroccan Arabic (Arabizi)	Latin/Numerals	800	0.979	0.990	0.984
ara_Arab	Modern Standard Arabic	Arabic	750	0.977	0.835	0.900
fra_Latn	French	Latin	800	0.995	0.998	0.996
eng_Latn	English	Latin	800	0.989	0.995	0.992
spa_Latn	Spanish	Latin	800	0.995	0.995	0.995
zho_Hani	Chinese	Hanzi	800	1.000	1.000	1.000

Kabyle precision is 0.900, up from 0.626 in Dihya-5M. Dihya-5M traded precision for recall, sending 37% of its Kabyle predictions to sibling varieties. Dihya-22M resolves that tension: Kabyle precision reaches 0.900 while maintaining 0.925 recall.

Central Atlas Tamazight recall rises from 0.660 to 0.948, resolving the largest weakness of the 5M checkpoint.

Shawiya recall reaches 0.714 on 98 test rows, against 0.687 in Dihya-5M, with precision rising to 0.864. 447 training rows is what exists in machine-readable form.

Arabic dialect discrimination

Separating Algerian Darija (arq), Moroccan Darija (ary), and Modern Standard Arabic (ara) in Arabic script and Arabizi:

True \ Predicted	arq_Arab	arq_Arabizi	ary_Arab	ary_Arabizi	ara_Arab
arq_Arab (Algerian Arabic)	708 (89.8%)	0	73 (9.3%)	0	7 (0.9%)
arq_Arabizi (Algerian Arabizi)	0	205 (91.9%)	0	14 (6.3%)	0
ary_Arab (Moroccan Arabic)	166 (20.8%)	0	625 (78.1%)	0	8 (1.0%)
ary_Arabizi (Moroccan Arabizi)	0	6 (0.8%)	0	792 (99.0%)	0
ara_Arab (Modern Standard Arabic)	83 (11.1%)	0	41 (5.5%)	0	626 (83.5%)

Arabizi and Arabic script have zero cross-leakage (0 errors). Over 89.8% of Algerian Darija and 78.1% of Moroccan Darija are separated correctly; on shared Maghrebi idioms the conformal set emits ('arq_Arab', 'ary_Arab').

By (language, domain) cell

Reporting is on 34 active (language, domain) cells with support $\ge 20$:

	recall
median cell	0.957 (vs 0.921 in 5M)
cells at or above 0.90	21 of 34
cells at or above 0.70	31 of 34
cells at 1.000	2 — Tashelhit scripture, Chinese foreign

The lowest cells with support $\ge 20$:

cell	support	recall
shy_Latn/bench_gold	30	0.633
shy_Latn/everyday	44	0.636
ary_Arab/negative_dialect	135	0.637
zgh_Tfng/everyday	758	0.740
shi_Tfng/everyday	800	0.744

The Shawiya cells reflect the language's digital scarcity. ary_Arab/negative_dialect holds Moroccan colloquial text sharing vocabulary with western Algerian dialects. zgh_Tfng and shi_Tfng share everyday lexicon in standard Neo-Tifinagh orthography.

Conformal prediction sets

Calibrated by split conformal at target coverage 99.0%. Fitted on one half of the development set (3,551 sentences) and coverage measured on the other (3,551 held-out sentences):

fitted on	3,551 dev sentences
coverage measured on	3,551 held-out dev sentences
target coverage	99.0%
empirical coverage	98.28%
quantile q̂	0.9526
mean set size	1.259 labels (across 22 classes)
singletons	79.86% (2,836 of 3,551)

Usage

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained(
    "agbalu/Dihya-22M", trust_remote_code=True
)

for row in model.identify([
    "Azul fell-awen, amek tettilim ass-a?",
    "Mayemmi ur teccid ticemrarin nnecm a gma?",
    "Bonjour tout le monde, comment allez-vous?",
    "ⵜⴰⵇⴱⴰⵢⵍⵉⵜ ⴷ ⵜⵓⵜⵍⴰⵢⵜ ⵜⴰⵢⴻⵎⵎⴰⵜ ⵏⵏⴻⵖ.",
    "Hello everyone, how are you doing today?",
    "صباح الخير كيداير لاباس عليك",
    "واش راك خويا لاباس؟",
]):
    print(row["language"], round(row["confidence"], 4), row["prediction_set"])
# kab_Latn 0.5295 ('kab_Latn', 'shy_Latn')
# rif_Latn 0.5485 ('kab_Latn', 'rif_Latn')
# fra_Latn 0.9444 ('fra_Latn',)
# kab_Tfng 0.9265 ('kab_Tfng',)
# eng_Latn 0.9498 ('eng_Latn',)
# ary_Arab 0.8180 ('arq_Arab', 'ary_Arab')
# arq_Arab 0.4889 ('arq_Arab', 'ary_Arab')

identify applies the conformal quantile that the numbers above were measured with; it travels in config.json. The plain forward returns raw cosine logits.

The tokenizer is the 256 UTF-8 byte values and is only needed for batching:

from transformers import AutoModelForSequenceClassification, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("agbalu/Dihya-22M")
model = AutoModelForSequenceClassification.from_pretrained(
    "agbalu/Dihya-22M", trust_remote_code=True
)
batch = tokenizer(["Azul fell-awen", "صباح الخير"], return_tensors="pt", padding=True, truncation=True)
logits = model(**batch).logits

Prediction sets widen where text is genuinely ambiguous. The greeting above returns ('kab_Latn', 'shy_Latn') — Shawiya and Kabyle share that greeting and neither carries a marker separating them. A forced singleton there would be an arbitrary guess.

Intended use

Screening and routing text in corpus pipelines: tagging web scrapes and OCR output across Latin, Tifinagh, and Arabic scripts, discarding non-target lines, and routing input to variety-specific models. Use the prediction set rather than the argmax where a wrong label would corrupt a corpus.

Not suitable for distinguishing sub-varieties inside any of the twenty-two classes, single isolated words lacking morphological context, or attributing authors on the basis of dialect.

Architecture

Parameters, trained	22,198,528
Parameters, published	21,999,616
Encoder layers	10 (vs 6 in 5M)
Hidden / feed-forward	384 / 1,344 (SwiGLU) (vs 256 / 704 in 5M)
Attention heads	6 × 64, rotary positions
Conv stem	four parallel depthwise-separable 1D convolutions, kernels 3, 5, 7, 11
Pooling	attentive (learned score per position)
Head	cosine, scale 28.0
Vocabulary	258 (256 UTF-8 bytes, [PAD], [UNK])
Maximum input	512 bytes (vs 256 in 5M)

The published model excludes the auxiliary script classification head, binary filter head, and contrastive projection head (198,912 parameters), which shape representations during training and are never read at inference.

Four kernel widths because evidence sits at four physical scales: grapheme clusters and diacritics ($k=3$), inflectional affixes ($k=5$), clitic chains and annexation markers ($k=7$), and word stems ($k=11$).

Training

training rows	113,974 across 36 (language, domain) cells
dev rows	7,145
schedule	8 epochs, 4,752 steps; batch 16 cells × 12 rows
published checkpoint	step 2,072, selected on dev selection score
dev score there	0.9157 macro-F1, 0.6364 worst-group recall
training time	24.8 minutes at 532 seq/s

Five objectives address long-tail imbalance, domain collapse, and multi-script structure:

1. LDAM margins for the long tail — $m_k = C / n_k^{1/4}$ (Cao et al., NeurIPS 2019).
2. Convex Group-DRO — exponentiated-gradient weights over the active (language, domain) groups (Sagawa et al., ICLR 2020) with convex normalization.
3. Domain-aware supervised contrastive learning — negatives mined from a different language in the same domain first (ConLID) over a 4,096-entry memory bank.
4. Multi-task auxiliary heads — script classification (5 scripts) and binary Amazigh filtering.
5. Deferred schedule — margins and group weights are disabled for the first half of training (steps 0–2,376).

Selection is on the mean of dev macro-F1 and dev worst-group recall, never on loss.

Limitations

Dialect continuum ambiguity: Short sentences of shared vocabulary carrying no variety-specific morphology cannot be deterministically resolved to a single class. The conformal prediction set captures this ambiguity directly by returning a multi-label set.

Shawiya data scarcity: 447 training rows and 98 test rows represent the entirety of machine-readable Shawiya text currently available; reported metrics reflect these sample size bounds.

Maghrebi Arabic dialect overlap: Algerian and Moroccan Arabic share substantial colloquial syntax and lexicon, leading to cross-dialect conformal sets on shared idioms.

Input length: Evaluated on sentence-level text; single isolated words lack sufficient phonotactic context for reliable attribution.

Code-switching: Outputs sentence-level classifications; token-level language diarization is not supported.

Files

file	description
model.safetensors	88.0 MB, 21,999,616 parameters
config.json	architecture, 22 classes, prior shift and conformal quantile
modeling_dihya.py, configuration_dihya.py	architecture, transformers and torch only
tokenizer.json, tokenizer_config.json	the 258-entry byte vocabulary
calibration.json	the calibration, in full
test-report.json	per class and per (language, domain) cell

No optimiser state, no scheduler state, no RNG. Training cannot be resumed from these files.

The name

Dihya (also called al-Kāhina, died c. 703 CE) was an Amazigh military leader from the Aurès Mountains of what is now eastern Algeria — the Shawiya homeland — who led the Berber resistance. The naming is homage and implies no endorsement by anyone.

Citation

@software{agbalu_dihya_22m_2026,
  title  = {Dihya-22M: Pan-Amazigh, North African Arabic, and World Language Identification},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/agbalu/Dihya-22M},
  note   = {22M parameters; 91.90\% accuracy over 22 classes on 14,247 held-out sentences}
}

Licence

Apache-2.0 on the weights and the code. The training corpus is assembled from sources under mixed licences; see resources/sibling_registry.yaml and resources/corpus_registry.yaml for per-source terms.

Part of AƔBALU, a Kabyle and Amazigh corpus and model collection.