Jisr-MT-50M-AllDialects

One 49M model, English → 13 Arabic dialects + MSA, selected by tag.

Trained to test whether dialects share enough structure to help each other. They largely do — it beats the single-dialect models on 5 of 9 comparable dialects and ties 2 more, at 1/13th the storage.

Results

UBC-NLP/alexandria test splits, beam 4. "single" is the dedicated 49M model for that dialect.

dialect	tag	single chrF++	this model	Δ chrF++	Δ BLEU
Libyan	>>ayl<<	24.17	31.79	+7.62	+6.29
Palestinian	>>ajp-ps<<	28.91	33.89	+4.98	+4.87
Syrian	>>ajp<<	34.68	37.92	+3.24	+3.06
Saudi	>>ars<<	34.50	35.44	+0.94	+1.18
Yemeni	>>acq<<	30.12	30.83	+0.71	+0.98
Sudanese	>>apd<<	34.11	34.60	+0.49	−0.37
Tunisian	>>aeb<<	30.58	30.83	+0.25	−0.42
Lebanese	>>apc<<	30.96	28.40	−2.56	−2.89
Moroccan	>>ary<<	32.23	30.50	−1.73	−3.66
Egyptian	>>arz<<	38.58¹	37.60	−0.98	−1.32

¹ vs Jisr-MT-50M-Masri-v2, the combined-corpus model.

Sudanese and Tunisian flip sign between the two metrics, so they are genuine ties. The honest tally is 5 clear wins, 3 clear losses, 2 ties.

The pattern is data-size dependent. Gains concentrate in small-corpus dialects — Libyan (12,850 pairs, +7.62) and Palestinian (29,283, +4.98) — and losses in the large ones: Moroccan (155,939) and Egyptian (which also has a second corpus). Dialects with enough data of their own are diluted by sharing capacity; dialects without enough are rescued by it.

MSA retention — the best in the family

model	BLEU	chrF++
un-finetuned base	30.77	51.09
this model	28.50	49.46
Jisr-MT-50M-Masri-v2	28.21	49.27

It saw 13 dialects' worth of MSA replay rather than one, so >>ara<< received proportionally more gradient.

Compared with a frontier model

gpt-5.1 on the same benchmark (200 turns per dialect, prompted to write dialect not MSA):

dialect	this model	gpt-5.1	Δ	% of gpt-5.1
Moroccan	30.50	20.86	+9.64	146%
Libyan	31.79	34.67	-2.88	92%
Sudanese	34.60	40.14	-5.54	86%
Palestinian	33.89	39.87	-5.98	85%
Tunisian	30.83	38.64	-7.81	80%
Egyptian	37.60	45.59	-7.99	82%
Syrian	37.92	46.16	-8.24	82%
Yemeni	30.83	39.32	-8.49	78%
Lebanese	28.40	37.19	-8.79	76%
Saudi	35.44	48.74	-13.30	73%

gpt-5.1 leads on 9 of 10, by 2.9–13.3 chrF++, averaging 39.21 against this model's ~34. The exception is Moroccan, where gpt-5.1 collapses to 20.86 — less than half its Saudi score — and this model leads by 9.64. See Jisr-MT-50M-Moroccan for that result.

gpt-5.1 also gains +2.3 to +4.8 chrF++ from orthographic normalisation while Jisr models gain ~+1.6, so the raw gaps understate it: trained in-domain, these models already match the references' spelling conventions.

Usage

from transformers import AutoTokenizer, MarianMTModel
tok = AutoTokenizer.from_pretrained("oddadmix/Jisr-MT-50M-AllDialects")
model = MarianMTModel.from_pretrained("oddadmix/Jisr-MT-50M-AllDialects")

def tr(text, tag):
    return tok.decode(model.generate(**tok(f"{tag} {text}", return_tensors="pt"),
                                     num_beams=4)[0], skip_special_tokens=True)

tr("Where are you going?", ">>arz<<")   # Egyptian
tr("Where are you going?", ">>ary<<")   # Moroccan
tr("Where are you going?", ">>ara<<")   # MSA

tag	dialect		tag	dialect
>>arz<<	Egyptian		>>ars<<	Saudi
>>ary<<	Moroccan		>>afb<<	Bahraini (Gulf)
>>arq<<	Algerian		>>apd<<	Sudanese
>>aeb<<	Tunisian		>>acq<<	Yemeni
>>ayl<<	Libyan		>>acm<<	Iraqi
>>apc<<	Lebanese		>>ajp-ps<<	Palestinian
>>ajp<<	Syrian		>>ara<<	MSA

The tag is required — the model cannot infer the target dialect. English → Arabic only.

Note >>apc<< (Lebanese) and >>ajp<< (Syrian) deviate from strict ISO 639-3, where both Levantine variants share apc. They were given distinct tags so the model can separate them.

Training

Base	Jisr-MT-50M-tatoeba (49M)
Data	1,458,160 rows across 13 dialects (Egyptian uses the combined corpus)
MSA replay	1×
Budget	1 epoch, bs 64, lr 3e-4 cosine, bf16, 57 min
Vocab	32,014 — 32k base plus 13 dialect tags, each seeded from >>ara<<

Limitations

• Higher degeneracy than the single models — up to 2.6% of outputs on Sudanese repeat themselves, vs ~1% for the dedicated models. Check outputs before shipping.
• Algerian, Bahraini and Iraqi are unvalidated. Alexandria has no config for them, so 3 of the 13 tags have no external benchmark at all.
• Dialect data is Gemma-translated ASR transcripts; for Egyptian, where a purpose-built parallel corpus exists, that corpus alone trains a better model.
• No standard orthography for any of these dialects — all chrF/BLEU here are depressed by that.
• English → Arabic only. 49M parameters.