NAMAA-Space/alexandria-backtranslated-pairs

AlexandriaX Back-Translated Dialectal Arabic Pairs

348,787 synthetic English→dialectal-Arabic training pairs over 14 Arabic varieties, built by the NAMAA Community for AlexandriaX-2026 Subtask 1 (ArabicNLP 2026 / EMNLP).

Each row is a ready-to-train instruction example: a prompt in the exact format the team's NileChat fine-tunes expect, and a response that is a real, human-written dialectal Arabic sentence harvested from public corpora. The English in the prompt is machine-generated — it is the nllb-200-distilled-600M back-translation of that Arabic sentence.

This augmentation did not work. Training NileChat-3B on it (84.7% synthetic against 15.3% real supervision) scored 22.77 spBLEU on the official AlexandriaX dev set against 23.54 for the same adapter trained on real data alone. No submitted system used it. It is released because the negative result is reusable, and because the underlying Arabic sentences are useful in their own right — but read Known limitations before building on it.

Rows	348,787 (single train split)
Direction	English (synthetic) → dialectal Arabic (authentic)
Varieties	14 — the 13 AlexandriaX countries plus Algerian (DZ)
Built from	358,483 cleaned dialectal sentences from public corpora
Back-translator	facebook/nllb-200-distilled-600M, fp16, greedy
Size	65 MB parquet, 214 MB in memory
Format	one prompt/response pair per row, prompt pre-rendered
License	CC-BY-NC-4.0 — non-commercial; the source corpora carry their own terms

Loading

from datasets import load_dataset

ds = load_dataset("NAMAA-Space/alexandria-backtranslated-pairs", split="train")
print(ds[0]["prompt"])
print(ds[0]["response"])

# Egyptian rows only
eg = ds.filter(lambda r: r["country"] == "EG")

The prompt field is already rendered in the format the sibling models were trained on, so it drops straight into a completion-only SFT run:

You are an expert translator. Translate the English sentence into EG.

### Metadata:
- Country: EG
- Domain: Everyday / Social
- Participants: Unknown Participants
- Speaker: A
- Speaker Direction: Unknown

### Conversation History:
No previous turns (Start of conversation).

### Sentence to Translate:
{NLLB back-translation of the Arabic sentence}

### Translation:

Fields

Field	Type	Description
prompt	string	Fully rendered instruction prompt. The metadata block is a placeholder: Domain is always Everyday / Social, Participants Unknown Participants, Speaker A, Speaker Direction Unknown, and the history is always No previous turns. Only Country carries real information.
response	string	The authentic dialectal Arabic sentence — the training target. Mean 84 characters, median 53, capped at 1,000.
country	string	ISO-2 country code of the variety (14 values, includes DZ).
conv_id	string	Synthetic id, BT-{CC}-{hash}. One row per id — these are not conversations.
turn_order	int64	Always 1. Present only for schema compatibility with the official task data.

Composition

Rows per variety:

Country	Rows	Share	In the AlexandriaX task?
SA Saudi	50,000	14.3%	yes (train + dev + test)
SY Syrian	50,000	14.3%	yes (train + dev + test)
MA Moroccan	46,286	13.3%	yes (train + dev + test)
PS Palestinian	30,538	8.8%	yes (train + dev + test)
DZ Algerian	30,078	8.6%	no — not a task variety
EG Egyptian	28,197	8.1%	yes (train + dev + test)
LB Lebanese	23,467	6.7%	yes (train + dev + test)
JO Jordanian	21,185	6.1%	yes (train + dev + test)
TN Tunisian	15,204	4.4%	yes (train + dev + test)
OM Omani	13,812	4.0%	yes (train + dev + test)
LY Libyan	13,587	3.9%	yes — test-only, so this is its only training signal
SD Sudanese	12,251	3.5%	yes — test-only, so this is its only training signal
YE Yemeni	12,031	3.4%	yes (train + dev + test)
MR Mauritanian	2,151	0.6%	yes (train + dev + test)

Saudi and Syrian are clipped at 50,000; Mauritanian Hassaniya is the scarcest variety here by an order of magnitude, exactly as it is in the task data — back-translation did not fix the long tail.

How it was built

The full pipeline is in this repo as build_backtranslated_pairs.py, extracted from the original Colab notebook (AlexandriaX_NB1_Preprocessing.ipynb).

1. Harvest. A monolingual corpus was assembled from public dialectal Arabic sources: IADD, QADI, NADI-2021, MADAR, and Egyptian, Saudi, Moroccan, Levantine, Libyan, Sudanese and Yemeni web and transcript collections. Ingestion is audio-free — for speech corpora only the transcript column is read, by column projection.
2. Clean, aggressively, because this is target-side text rather than parallel data. Result: 358,483 sentences across 14 varieties.
3. Back-translate each sentence to English, then invert the pair.
4. Render as an instruction prompt with placeholder metadata. 348,787 rows survived — the shortfall against 358,483 is the 50,000-per-dialect cap biting on SA (52,411 available) and SY (57,283).

Cleaning parameters, exactly

Step	Setting
HTML unescape	yes
URLs	`https?://\S+
@-mentions	@\w+ → space
Hashtags	#\w+ → space (the whole tag, not just #)
Timestamps	\b\d{1,2}:\d{2}(:\d{2})?\s?[ap]m?\b → space
Emoji	stripped (5 Unicode blocks)
Tatweel ـ	removed
Character runs	3+ repeats of any character collapsed to 1
Diacritics (tashkeel)	kept (REMOVE_DIACRITICS = False)
Latin letters	deleted (REMOVE_LATIN_TOKENS = True) — see the bias note below
Punctuation	_ / \ ~ @ # $ % ^ & * ( ) - + = [ ] { } | ; : " ' < > , . ? → space
Unicode	NFKC normalisation, whitespace collapsed
Length filter	3 ≤ characters ≤ 1,000, ≥ 1 word
Arabic-ratio filter	≥ 0.50 — but applied after Latin removal, so in practice it only drops rows left with no Arabic at all
Near-duplicates	dropped per dialect on a key of (tashkeel-stripped, whitespace-removed, lowercased) text

Back-translation parameters, exactly

Setting	Value
Model	facebook/nllb-200-distilled-600M (the distilled 600M, not the 1.3B used for the forward-direction fine-tune)
Precision	float16, single CUDA device
Direction	dialect → English, forced_bos_token_id = eng_Latn
Decoding	greedy — no beams, no sampling
max_new_tokens	128
Source truncation	128 tokens
Batch size	256 (sized for an L4)
Cap per variety	50,000 (BT_PER_DIALECT)
conv_id	BT-{CC}-{md5(text)[:8]} — also the resume key
Output	JSONL, appended and flushed per row so an interrupted run resumes

Per-variety source codes for the dialect→English pass. Note these differ from the codes used to fine-tune NLLB in the forward direction: here LY and SD get dedicated codes, while OM and MR fall back to MSA.

EG	MA	TN	LY	SD	SA	YE	OM	JO	LB	SY	PS	MR	DZ / other
arz	ary	aeb	ayl	apd	ars	acq	arb	ajp	apc	apc	ajp	arb	arb

(all _Arab script)

Measured effect

Same base model, same QLoRA hyperparameters, one epoch, greedy decoding; scored country-macro on the official AlexandriaX dev set (12,250 turns, 11 countries) and on an internal hold-out of 3,350 turns carved from the training conversations.

Training data	dev spBLEU	dev chrF++	hold-out spBLEU	hold-out chrF++
63,130 real turns, no context	23.54	39.68	24.84	40.36
63,130 real turns, with context	22.87	39.11	23.37	39.35
+ 348,787 rows of this dataset, with context	22.77	38.71	24.36	40.16

The two evaluation sets disagree: on the internal hold-out the augmentation gains +0.99 spBLEU over the context baseline, on the official dev set it loses 0.10. The hold-out shares conversations, speakers and phrasing with the training data, so it flatters anything that adds volume. Per country on official dev, the augmentation helped where real data is scarce (TN +1.42, OM +0.87, EG +0.34) and hurt where it is plentiful (SY −0.93, PS −0.83, SA −0.84, MA −0.82).

Trained checkpoint: alexandriax-nilechat-ctx-aux.

Known limitations

• The English side is machine-generated by a 600M distilled translator and is not human-verified. Some of it is visibly disfluent or semantically drifted; models trained on it learn to translate NLLB-distilled English, which is not the register of the task's hand-written dialogue turns.
• Latin script is deleted from the targets, not merely filtered. REMOVE_LATIN_TOKENS=True strips every A-Za-z character before back-translation, so no row contains a single Latin token. AlexandriaX gold does the opposite: 33.8% of Moroccan and 39.1% of Tunisian gold turns carry Latin-script French or English borrowings. Training on this data therefore pushes a model away from the reference style of exactly the varieties the augmentation was meant to help — and it mutilates code-switched sentences rather than dropping them, since the Arabic-ratio filter runs afterwards and no longer sees any Latin.
• Punctuation is stripped from the targets but not from the English prompts, so a model trained here learns to drop sentence-final periods, commas and question marks against punctuated references. (ASCII ! and the Arabic ؟ happen to survive the filter, which makes the loss inconsistent rather than cleanly systematic.)
• No conversation structure. Every row is a first turn with empty history and placeholder metadata, so this data cannot teach context-aware behaviour — which is what the task rewards.
• Metadata fields are fake apart from country. Do not treat Domain, Speaker or Speaker Direction as labels.
• DZ (Algerian) is out of task scope — 8.6% of the rows target a variety AlexandriaX never evaluates.
• Variety labels are inherited, not verified. They come from the source corpora's own country labels; dialect identification noise in those corpora propagates here, and country-level labels flatten real sub-dialect variation.
• Provenance is coarse. Per-row source attribution was not retained, so a row cannot be traced back to IADD vs QADI vs a web dump. Redistribution terms of the underlying corpora differ; check them before any use beyond research.
• Non-commercial. Web and corpus provenance plus the NLLB licence make this research-only.

---

The shared task

AlexandriaX-2026 (ArabicNLP 2026 / EMNLP) — Context-Aware Dialectal Arabic MT and MT Evaluation. This model was built for Subtask 1: Context-Aware English-to-Dialectal Arabic Dialogue Translation.

Given one English dialogue turn together with its conversation history and metadata — target country/dialect, domain, participant roles, speaker, and speaker→addressee gender direction — the system must produce the turn in the requested country's spoken Arabic, preserving meaning while adapting lexical, morphological, pragmatic and sociolinguistic choices to that variety.

Two tracks: constrained (provided data only, ≤5B parameters) and unconstrained (any external data or model). Ranking is by spBLEU (primary) and chrF++ (secondary), each macro-averaged over countries.

Official data (UBC-NLP/alexandria)

Split sizes in turns, as published by the organisers:

Split	EG	JO	LB	LY	MA	MR	OM	PS	SA	SD	SY	TN	YE	Total
train	3,108	5,501	8,906	0	2,573	5,515	6,280	14,933	8,470	0	6,071	2,034	3,089	66,480
dev	1,113	1,113	1,118	0	1,110	1,114	1,109	1,110	1,110	0	1,119	1,116	1,118	12,250
public test	1,118	1,107	1,106	1,109	1,115	1,112	1,118	1,109	1,113	1,106	1,114	1,109	1,106	14,442
private (blind) test	1,113	1,109	1,110	1,309	1,111	1,119	1,107	1,111	1,114	915	1,114	1,114	1,113	14,459

Libyan (LY) and Sudanese (SD) appear only at test time — they are zero-shot for every system trained on this data.

Conversation-level counts: 21,146 train / 3,963 dev / 4,706 public-test conversations; mean 3.13 turns per conversation (range 1–5). Mean length 102 characters of English source, 74 characters of dialectal target.

Dialects (13 countries). Egyptian, Jordanian, Lebanese, Libyan, Moroccan, Mauritanian, Omani, Palestinian, Saudi, Sudanese, Syrian, Tunisian, Yemeni. Labels are country + sub-dialect, and several countries carry more than one: Palestinian 10 (Nabulsi and Albira urban, plus Falahi varieties of Surif, Kobar, Noba, Ni'lin, Shuqba, Aboud, Silwad, Ramallah), Omani 5 (Suri, Rustaqi, Al-Wafi, Ibri, Seebi), Saudi 3 (Southern, Hijazi, Khaleeji), Yemeni 3 (Taiz, San'ani, Central), Syrian 2 (Levantine Standard, Homsi). The remaining countries carry one label each (e.g. Egyptian Arabic (Cairene), Moroccan Standard Darija, Mauritanian Hassaniya, Libyan Arabic (Misrati/Central)).

Domains (11, near-uniform). Agriculture and farming, Commerce and transactions, Construction and real estate, Education and academia, Energy and resources, Everyday and social, Healthcare and medical, Legal and financial, Logistics and transportation, Professional and workplace, Tourism and hospitality.

Speaker direction (turns, train+dev+public test): female→male 30,636 · male→female 30,203 · male→male 20,465 · female→female 11,868. The corpus carries 76 distinct translator IDs and 44 reviewer IDs.

Code-switching in the gold is strongly dialect-specific — the share of gold turns containing Latin characters runs from TN 39.1% / MA 33.8% / LB 18.1% down to SY 1.2% / YE 0.8%. Systems that normalise every borrowing into Arabic script are penalised hardest on Maghrebi references (see Known limitations).

Evaluation protocol

• spBLEU — sacrebleu.BLEU(tokenize="flores200"), corpus-level per country, then averaged over countries.
• chrF++ — sacrebleu.CHRF(word_order=2), same averaging.
• Decoding is turn-by-turn: at turn n the conversation history contains the system's own previous outputs, never the gold ones. (An early evaluation harness in this project leaked gold previous-turn Arabic into the prompt and inflated scores by ≈2.4 spBLEU; every number reported here comes from the corrected, self-conditioned harness.)

---

The collection

All released artefacts live in NAMAA at AlexandriaX-2026:

Repo	What it is
alexandriax-arat5v2-base	AraT5v2-base full fine-tune — best small fine-tune, 25.12 dev / 23.26 blind spBLEU
alexandriax-arat5v2-balanced	the same recipe on a temperature-rebalanced dialect mixture — void run, released for the post-mortem and the fixed script
alexandriax-nilechat-lora	NileChat-3B QLoRA, context-free — best of the three NileChat variants, 23.54 dev spBLEU
alexandriax-nilechat-ctx-aux	NileChat-3B QLoRA, context + back-translation — the augmentation ablation, 22.77 dev spBLEU
alexandriax-nllb-1.3b-lora	NLLB-200-1.3B QLoRA with per-dialect language codes, 21.83 dev spBLEU
alexandriax-mt5-large-balanced	mT5-large on the rebalanced mixture — partial run (2,500/31,568 steps), never evaluated
alexandria-backtranslated-pairs	348,787 synthetic EN→dialect pairs over 14 varieties

Every model repo above carries a single-file train_*.py reproduction script with the exact hyperparameters that produced its checkpoint; the dataset repo carries build_backtranslated_pairs.py.

Official task data: UBC-NLP/alexandria. Base models: UBC-NLP/AraT5v2-base-1024, UBC-NLP/NileChat-3B-Base, facebook/nllb-200-1.3B, google/mt5-large.

---

Team

NAMAA Community — Fatimah Emad Eldin (Cairo University) · Omer Nacar (Tuwaiq Academy) · Khloud Al Jallad (Arab International University) · Mona Abdelazim (Ain Shams University).

Citation

Coming soon. The NAMAA system-description paper for AlexandriaX-2026 is under review for the ArabicNLP 2026 (EMNLP) proceedings; this card will be updated with the final ACL Anthology reference and DOI when the proceedings are published. Until then, please cite as:

@inproceedings{namaa-alexandriax-2026,
  title     = {{NAMAA} Community at {AlexandriaX-2026}: Prompting, Fine-Tuning and Agreement
               Voting for Dialectal Arabic Translation and Evaluation},
  author    = {Emad Eldin, Fatimah and Nacar, Omer and Al Jallad, Khloud and Abdelazim, Mona},
  booktitle = {Proceedings of the Fourth Arabic Natural Language Processing Conference
               (ArabicNLP 2026)},
  year      = {2026},
  note      = {To appear. Citation coming soon.}
}

Please also cite the shared task and the base model:

@inproceedings{alexandriax2026,
  title     = {{AlexandriaX-2026} Shared Task: Context-Aware Dialectal Arabic Machine
               Translation and MT Evaluation},
  author    = {El Mekki, Abdellah and Elmadany, AbdelRahim A. and Magdy, Samar M. and
               Ezzini, Saad and El-Haj, Mo and Jarrar, Mustafa and El-Beltagy, Samhaa and
               Abbas, Mourad and Zaraket, Fadi and Al Mandhari, Salim and Alyafeai, Zaid and
               Ghanem, Bernard and Abdul-Mageed, Muhammad},
  booktitle = {Proceedings of the Fourth Arabic Natural Language Processing Conference
               (ArabicNLP 2026)},
  year      = {2026},
  note      = {Overview paper. Citation coming soon.}
}

Acknowledgements

Thanks to the AlexandriaX-2026 organisers for the data, the evaluation infrastructure and their responsiveness during the evaluation phases.