alkhder at AlexandriaX-2026: Context-Aware English-to-Dialectal Arabic Dialogue Translation with QLoRA-Fine-Tuned NileChat-3B

Hasan Alkhder; Maher Pasa; Maria Hanini; Amro Najjar; Ahmet Zengin

cs.CLOct 15, 2026

alkhder at AlexandriaX-2026: Context-Aware English-to-Dialectal Arabic Dialogue Translation with QLoRA-Fine-Tuned NileChat-3B

Authors: Hasan Alkhder, Maher Pasa, Maria Hanini, Amro Najjar, Ahmet Zengin

Organizations: Sakarya University, Türkiye · Altınbaş University, Türkiye · Imperial College London · Luxembourg Institute of Science and Technology

Abstract

This paper describes team Alkhder's submission to Subtask 1 of the AlexandriaX-2026 shared task, which asks systems to translate English dialogue turns into thirteen Arabic dialects. We fine-tune NileChat-3B-Base with QLoRA on the Alexandria parallel dialogue corpus, training one set of adapter weights for all dialects. Each turn is translated conditioned on the target country and dialect, the domain, the participant roles, the speaker-addressee gender direction, and the preceding dialogue context. At inference, conversations are translated autoregressively and the model is fed its own previous outputs back as context without using any gold target-side test information. Under the constrained track, using only Alexandria data, it trains in about one hour and translates the private test set in under two hours. The system reaches a country-macro-averaged spBLEU of 20.36 and chrF++ of 36.64 on both tracks.

Figures & tables

Country	Available	Used (cap 2,500)
EG (Egypt)	4,221	2,500
JO (Jordan)	6,614	2,500
LB (Lebanon)	10,024	2,500
MA (Morocco)	3,683	2,500
MR (Mauritania)	6,629	2,500
OM (Oman)	7,389	2,500

Table 1: Turn-level training examples per country. Libya and Sudan have no training data.

Panel of Figure 1 — Figure 1: System overview.

Constrained track
#	Team	spBLEU	chrF++	Ratio
1	Thakaa	30.42	45.42	1.49
2	LahjaMT	28.30	43.81	1.55
3	FCDS	26.48	42.30	1.60
4	Rosetta	26.10	41.79	1.60
5	Namaa Community	23.26	39.03	1.68

Table 2: Official Subtask 1 results.

Country	Ref	Sys	Ref	Sys	Tok ratio
	Turns (%)		Tokens (%)
TN	47.5	32.3	6.88	3.97	0.58
MA	29.5	18.9	4.02	2.34	0.58
LB	21.2	10.4	2.32	1.12	0.48
EG	10.2	4.3	0.94	0.33	0.35
MR	7.0	1.3	0.74	0.11	0.15

Table 3: Latin-script code-switching on the matched public-test pair.

Context condition	spBLEU	chrF++
B: gold Arabic history	21.62	37.66
C: empty history	21.41	37.23
A: self-generated (submitted)	20.56	36.73

Table 4: Context ablation on 200 conversations per country (8,000 turns), scored against public-test references.

Country	A: self	B: gold	C: empty	B $-$ A	C $-$ A
SY	26.88	27.48	28.08	+0.60	+1.20
JO	25.95	27.91	27.07	+1.96	+1.12
TN	25.29	26.38	26.32	+1.09	+1.03
SA	24.78	26.84	26.02	+2.06	+1.24
EG	23.80	24.81	24.43	+1.01	+0.63
PS	21.02	22.74	22.60	+1.72	+1.58

Table 5: Per-country context ablation (macro spBLEU).

	Raw	Normalised	$\Delta$
spBLEU (macro)	20.80	22.36	+1.56
chrF++ (macro)	37.01	38.28	+1.27

Table 6: Effect of orthographic normalzation on the public-test evaluation.

Country	MSA $\Delta$ (system $-$ reference)
TN	$+0.35$
SD	$+0.34$
MR	$+0.18$
EG	$+0.16$
SY	$+0.09$
LY	$+0.09$

Table 7: MSA function-word rate difference.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Setting
Base model	UBC-NLP/NileChat-3B-Base ( $\approx$ 3B params; limit 5B)
Quantisation	4-bit NF4, double quantisation, bf16 compute (bitsandbytes)
LoRA	$r=16$ , $\alpha=32$ , dropout $0.05$ ; targets q/k/v/o_proj , gate/up/down_proj
Epochs	1
Learning rate	$2\mathrm{e}{-4}$ , cosine schedule, 20 warmup steps
Batch size	8 per device $\times$ 2 grad. accumulation (effective 16); 1,719 steps

Appendix

Table 8: System configuration

Ctry	spBLEU	chrF++	$\Delta$
	Public test		Norm.
SY	26.31	42.88	$+0.73$
JO	26.00	42.14	$+1.33$
TN	25.11	39.37	$+0.79$
SA	24.74	40.92	$+2.68$
EG	23.41	38.33	$+2.08$

Appendix

Table 9: Per-country results on the public test split. Norm. is the spBLEU gain from orthographic normalization. ∗ zero-shot varieties.

Category	Turns
Error-free	23
Content / meaning error	19
Dialect–MSA confusion	11
Morphological agreement	8
Unnatural, content acceptable	6
Orthographic error	2

Appendix

Table 10: Annotation counts over 60 sampled turns.

Machine Translation

Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.

See all 254 papers

Rosetta at AlexandriaX-2026: LoRA-Adapted NileChat for Context-Aware Dialectal Arabic Dialogue Translation
Sep 9, 2026Nada Esmaeil, Fathima Rena, Sibi Subhash +4Arabic NLP Low-Resource MT
Tanta University, Egypt · Alliance University, India · Robusta Studio, Egypt +2
AlexandriaX 2026: The First Shared Task on Dialectal Arabic Machine Translation
Sep 19, 2026Abdellah El Mekki, AbdelRahim A. Elmadany, Samar M. Magdy +6Context-Aware MT Machine Translation
The University of British Columbia · King Fahd University of Petroleum and Minerals · Lancaster University +4
Aladdin-FTI @ AMIYA Three Wishes for Arabic NLP: Fidelity, Diglossia, and Multidialectal Generation
Feb 18, 2026Jonathan Mutal, Perla Al Almaoui, Simon Hengchen +1Arabic NLP Arabic Dialects
Faculté de traduction et d’interprétation, Université de Genève

Sep 9, 2026J/K move · Enter open · S save

Rosetta at AlexandriaX-2026: LoRA-Adapted NileChat for Context-Aware Dialectal Arabic Dialogue Translation

Nada Esmaeil, Fathima Rena, Sibi Subhash, Osama Elgendy+3

Tanta University, Egypt · Alliance University, India · Robusta Studio, Egypt+2

Arabic NLP Low-Resource MT Context-Aware MT Machine Translation Arabic Dialects

Open paper AskPDF arXiv

This paper describes the Rosetta system for Subtask 1 (Context-Aware English-to-Dialectal Arabic Dialogue Translation) of the AlexandriaX shared task, participating in both constrained and unconstrained tracks. The approach fine-tunes a LoRA adapter on NileChat-3B using structured system/user prompts that condition generation on dialect and dialogue context. For the unconstrained track, the adapter is additionally pretrained on MADAR and PADIC. Rosetta ranked 4th in the constrained track (spBLEU 26.10) and 5th in the unconstrained track (spBLEU 25.09). The experimental results demonstrate that external pretraining helps only two of thirteen dialects while slightly hurting overall performance, suggesting negative transfer.

Abstract

Machine Translation

Explore similar work

Rosetta at AlexandriaX-2026: LoRA-Adapted NileChat for Context-Aware Dialectal Arabic Dialogue Translation

AlexandriaX 2026: The First Shared Task on Dialectal Arabic Machine Translation

Aladdin-FTI @ AMIYA Three Wishes for Arabic NLP: Fidelity, Diglossia, and Multidialectal Generation