Alexandria covers 13 Arab countries, 11 domains, and 107K community-driven samples.
Alexandria is a multi-domain English↔Dialectal Arabic machine translation dataset designed for culturally inclusive, dialect-aware NLP and LLM evaluation. It pairs English multi-turn conversations with human-translated dialectal Arabic from 13 Arab countries, enriched with sub-dialect metadata (based on city-level information), domain labels, persona roles, and speaker→addressee gender configurations. Alexandria is built to support both training and benchmarking for Arabic machine translation, dialectal Arabic generation, conversation-aware translation, and LLM evaluation across regional Arabic varieties.
This dataset could be useful for developpers and researchers looking for dialectal Arabic machine translation, English Arabic translation dataset, Arabic dialect benchmark, multi-domain Arabic MT corpus, and city-level Arabic dialect dataset.
Arabic is highly diglossic: most day-to-day communication happens in regional dialects rather than Modern Standard Arabic (MSA). Alexandria was created to narrow that gap by providing a large-scale, community-driven, human-translated resource that captures real dialect diversity across countries, cities, domains, and gendered conversational settings.
Alexandria contains 34,488 multi-turn conversations and approximately 107K total turns. The dataset covers 13 Arab countries and 11 socially important domains, including healthcare, education, agriculture, commerce, legal/financial services, logistics, tourism, and workplace communication. Unlike earlier resources that are limited to sentence-level translation or coarse regional labels, Alexandria is conversation-based, city-anchored, and metadata-rich, making it especially useful for evaluating dialect authenticity, contextual translation quality, and culturally grounded language understanding.
The dataset serves two main purposes:
from datasets import load_dataset
# Load a specific country subset (e.g., 'MA' for Morocco, 'EG' for Egypt) with a specific split
train_data = load_dataset("UBC-NLP/alexandria", name="MA", split="train")
test_data = load_dataset("UBC-NLP/alexandria", name="MA", split="test")
# View the first parallel turn of the first conversation from the train set
first_conv = train_data[0]
eng_turn = first_conv['english_conversation'][0]
dialect_turn = first_conv['dialectal_conversation'][0]
print(f"English: {eng_turn['text']}")
print(f"Dialect: {dialect_turn['text']}")
Alexandria is intended for:
Alexandria is a parallel, turn-aligned, multi-turn conversation dataset. Conversations are generated in English, then translated and revised by native speakers of local Arabic dialects.
Each example is part of a conversation and is expected to include, at minimum, content equivalent to:
Countries / dialect groups:
Domains:
Alexandria was created to address a major gap in Arabic NLP: most Arabic MT systems still perform best on MSA or on limited dialect benchmarks, even though real-world communication across the Arab world is predominantly dialectal. Existing resources often have one or more of the following limitations:
Alexandria was designed to provide a more realistic benchmark and training resource for dialect-sensitive, culturally grounded Arabic MT.
The dataset combines LLM-generated English source conversations with human translation and revision by native speakers of local dialects.
The dataset creation pipeline has three major phases:
English source generation
Human translation into dialectal Arabic
Peer revision
The English source conversations in Alexandria were constrained to exclude personally identifiable information (PII). Automated and human screening steps were also used to detect privacy issues, cultural mismatches, and unrealistic content before translation. Contributors were given explicit guidance around local norms, data privacy, and informed consent.
That said, users should still treat the dataset as human language data and review it before deploying models in sensitive or user-facing settings.
If you use Alexandria, please cite the paper introducing the dataset.
BibTeX:
@inproceedings{el-mekki-etal-2026-alexandria,
title = "Alexandria: A Multi-Domain Dialectal {A}rabic Machine Translation Dataset for Culturally Inclusive and Linguistically Diverse {LLM}s",
author = "EL Mekki, Abdellah and
Magdy, Samar M. and
Atou, Houdaifa and
AbuHweidi, Ruwa and
Qawasmeh, Baraah and
Nacar, Omer and
Al-hibiri, Thikra and
Saadie, Razan and
Alsayadi, Hamzah A. and
Hammouda, Nadia Ghezaiel and
Alkhazimi, Alshima Mohammed and
Hamod, Aya and
Al-Ghafri, Al-Yas Yaqoob and
El-Sayed, Wesam and
al Sharji, Asila Ismail and
Ballout, Mohamad and
Belfathi, Anas and
Ghaddar, Karim and
Sibaee, Serry and
Aoun, Alaa and
Aseri, Aeej Mohammed and
Abureesh, Lina and
Bashiti, Ahlam and
Yousef, Majdal and
Hafiz, Abdulaziz and
Mohamed, Yehdih and
Hamedtou, Emira and
Emehah, Brakehe and
Alhamouri, Rahaf and
Nafea, Youssef and
El Aatar, Aya and
Al-Dhabyani, Walid and
Hamed, Emhemed S. and
Shatnawi, Sara and
Alwajih, Fakhraddin and
Elkhidir, Khalid and
Alasmari, Ashwag and
Gerrio, Abdurrahman and
Alshahri, Omar Said and
Elmadany, AbdelRahim A. and
Berrada, Ismail and
Al-kathiri, Amir Azad Adli and
Zaraket, Fadi and
Jarrar, Mustafa and
EL Hadj, Yahya Mohamed and
Alhuzali, Hassan and
Abdul-Mageed, Muhammad",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.1503/",
pages = "32567--32592",
ISBN = "979-8-89176-390-6"
}
[email protected][email protected]Alexandria covers 13 Arab countries, 11 domains, and 107K community-driven samples.
Alexandria is a multi-domain English↔Dialectal Arabic machine translation dataset designed for culturally inclusive, dialect-aware NLP and LLM evaluation. It pairs English multi-turn conversations with human-translated dialectal Arabic from 13 Arab countries, enriched with sub-dialect metadata (based on city-level information), domain labels, persona roles, and speaker→addressee gender configurations. Alexandria is built to support both training and benchmarking for Arabic machine translation, dialectal Arabic generation, conversation-aware translation, and LLM evaluation across regional Arabic varieties.
This dataset could be useful for developpers and researchers looking for dialectal Arabic machine translation, English Arabic translation dataset, Arabic dialect benchmark, multi-domain Arabic MT corpus, and city-level Arabic dialect dataset.
Arabic is highly diglossic: most day-to-day communication happens in regional dialects rather than Modern Standard Arabic (MSA). Alexandria was created to narrow that gap by providing a large-scale, community-driven, human-translated resource that captures real dialect diversity across countries, cities, domains, and gendered conversational settings.
Alexandria contains 34,488 multi-turn conversations and approximately 107K total turns. The dataset covers 13 Arab countries and 11 socially important domains, including healthcare, education, agriculture, commerce, legal/financial services, logistics, tourism, and workplace communication. Unlike earlier resources that are limited to sentence-level translation or coarse regional labels, Alexandria is conversation-based, city-anchored, and metadata-rich, making it especially useful for evaluating dialect authenticity, contextual translation quality, and culturally grounded language understanding.
The dataset serves two main purposes:
from datasets import load_dataset
# Load a specific country subset (e.g., 'MA' for Morocco, 'EG' for Egypt) with a specific split
train_data = load_dataset("UBC-NLP/alexandria", name="MA", split="train")
test_data = load_dataset("UBC-NLP/alexandria", name="MA", split="test")
# View the first parallel turn of the first conversation from the train set
first_conv = train_data[0]
eng_turn = first_conv['english_conversation'][0]
dialect_turn = first_conv['dialectal_conversation'][0]
print(f"English: {eng_turn['text']}")
print(f"Dialect: {dialect_turn['text']}")
Alexandria is intended for:
Alexandria is a parallel, turn-aligned, multi-turn conversation dataset. Conversations are generated in English, then translated and revised by native speakers of local Arabic dialects.
Each example is part of a conversation and is expected to include, at minimum, content equivalent to:
Countries / dialect groups:
Domains:
Alexandria was created to address a major gap in Arabic NLP: most Arabic MT systems still perform best on MSA or on limited dialect benchmarks, even though real-world communication across the Arab world is predominantly dialectal. Existing resources often have one or more of the following limitations:
Alexandria was designed to provide a more realistic benchmark and training resource for dialect-sensitive, culturally grounded Arabic MT.
The dataset combines LLM-generated English source conversations with human translation and revision by native speakers of local dialects.
The dataset creation pipeline has three major phases:
English source generation
Human translation into dialectal Arabic
Peer revision
The English source conversations in Alexandria were constrained to exclude personally identifiable information (PII). Automated and human screening steps were also used to detect privacy issues, cultural mismatches, and unrealistic content before translation. Contributors were given explicit guidance around local norms, data privacy, and informed consent.
That said, users should still treat the dataset as human language data and review it before deploying models in sensitive or user-facing settings.
If you use Alexandria, please cite the paper introducing the dataset.
BibTeX:
@inproceedings{el-mekki-etal-2026-alexandria,
title = "Alexandria: A Multi-Domain Dialectal {A}rabic Machine Translation Dataset for Culturally Inclusive and Linguistically Diverse {LLM}s",
author = "EL Mekki, Abdellah and
Magdy, Samar M. and
Atou, Houdaifa and
AbuHweidi, Ruwa and
Qawasmeh, Baraah and
Nacar, Omer and
Al-hibiri, Thikra and
Saadie, Razan and
Alsayadi, Hamzah A. and
Hammouda, Nadia Ghezaiel and
Alkhazimi, Alshima Mohammed and
Hamod, Aya and
Al-Ghafri, Al-Yas Yaqoob and
El-Sayed, Wesam and
al Sharji, Asila Ismail and
Ballout, Mohamad and
Belfathi, Anas and
Ghaddar, Karim and
Sibaee, Serry and
Aoun, Alaa and
Aseri, Aeej Mohammed and
Abureesh, Lina and
Bashiti, Ahlam and
Yousef, Majdal and
Hafiz, Abdulaziz and
Mohamed, Yehdih and
Hamedtou, Emira and
Emehah, Brakehe and
Alhamouri, Rahaf and
Nafea, Youssef and
El Aatar, Aya and
Al-Dhabyani, Walid and
Hamed, Emhemed S. and
Shatnawi, Sara and
Alwajih, Fakhraddin and
Elkhidir, Khalid and
Alasmari, Ashwag and
Gerrio, Abdurrahman and
Alshahri, Omar Said and
Elmadany, AbdelRahim A. and
Berrada, Ismail and
Al-kathiri, Amir Azad Adli and
Zaraket, Fadi and
Jarrar, Mustafa and
EL Hadj, Yahya Mohamed and
Alhuzali, Hassan and
Abdul-Mageed, Muhammad",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.1503/",
pages = "32567--32592",
ISBN = "979-8-89176-390-6"
}
[email protected][email protected]