@inproceedings{wahid-etal-2026-dynamos,
title = "Dynamos at {A}r{G}uard Shared Tasks: Multimodal Transformers and Selective Ensembling for Binary and Fine-Grained {A}rabic Hateful Meme and Harmful Prompt Detection",
author = "Wahid, Sheikh Abdul and
Yamin, Muhammad and
Arshad, Ali",
editor = "Abdul-Mageed, Muhammad and
Elsayed, Tamer and
Bouamor, Houda and
El-Beltagy, Samhaa and
Al-Khalifa, Hend and
Abdelali, Ahmed and
Elmadany, AbdelRahim and
El Mekki, Abdellah and
Darwish, Kareem and
Alkhereyf, Sakhar and
Luqman, Hamzah and
Khalifa, Salam and
Magdy, Samar and
Alyafeai, Zaid and
Hamed, Injy and
Zaghouani, Wajdi and
Zitouni, Imed",
booktitle = "Proceedings of The Fourth {A}rabic Natural Language Processing Conference: Shared Tasks",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.arabicnlp-sharedtasks.46/",
pages = "371--378",
ISBN = "979-8-89176-497-2",
abstract = "For the ArGuard 2026 shared task on harmful content detection in Arabic memes (Task A) and LLM prompts (Task B), we fine-tune Arabic transformer encoders across all four subtasks. Memes additionally use a vision tower (ViT-B/16 for the binary subtask, CLIP ViT-L/14 for the fine-grained one) late-fused with MARBERTv2; ensembling choices and decision thresholds are tuned per subtask throughout. Two methodological choices account for much of our gains: selecting checkpoints by development macro-F1 rather than loss, and aligning the B2 training target with the evaluation distribution. Our largest single effect, however, was diagnostic. The B1 blind test set is approximately class-balanced while its training data is {\textasciitilde}80/20, so our dev-tuned threshold collapsed macro-F1 from 0.9557 on dev to 0.6345 on test. We recovered the test class balance algebraically from the aggregate metrics the platform reported to all participants; re-thresholding the same saved probabilities, with no retraining, restored 0.8628. Because this uses only aggregate leaderboard feedback (within the task rules), the B1 figure is not directly comparable to teams that did not use the channel and is not a generalisation estimate. This correction is also not general: on the weaker multi-label classifier (A2), stronger prior matching reduced performance in every setting we tested, so matching the test prior is not necessarily F1-optimal. We additionally report a catalogue of quantified negative results, including external-data augmentation with AraSafe. Final test macro-F1: A1 0.7869, A2 0.3463, B1 0.8628, B2 0.7622."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wahid-etal-2026-dynamos">
<titleInfo>
<title>Dynamos at ArGuard Shared Tasks: Multimodal Transformers and Selective Ensembling for Binary and Fine-Grained Arabic Hateful Meme and Harmful Prompt Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sheikh</namePart>
<namePart type="given">Abdul</namePart>
<namePart type="family">Wahid</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="family">Yamin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ali</namePart>
<namePart type="family">Arshad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of The Fourth Arabic Natural Language Processing Conference: Shared Tasks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="family">Abdul-Mageed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tamer</namePart>
<namePart type="family">Elsayed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Houda</namePart>
<namePart type="family">Bouamor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samhaa</namePart>
<namePart type="family">El-Beltagy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hend</namePart>
<namePart type="family">Al-Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ahmed</namePart>
<namePart type="family">Abdelali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">AbdelRahim</namePart>
<namePart type="family">Elmadany</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Abdellah</namePart>
<namePart type="family">El Mekki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kareem</namePart>
<namePart type="family">Darwish</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sakhar</namePart>
<namePart type="family">Alkhereyf</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hamzah</namePart>
<namePart type="family">Luqman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Salam</namePart>
<namePart type="family">Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samar</namePart>
<namePart type="family">Magdy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zaid</namePart>
<namePart type="family">Alyafeai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Injy</namePart>
<namePart type="family">Hamed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wajdi</namePart>
<namePart type="family">Zaghouani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Imed</namePart>
<namePart type="family">Zitouni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-497-2</identifier>
</relatedItem>
<abstract>For the ArGuard 2026 shared task on harmful content detection in Arabic memes (Task A) and LLM prompts (Task B), we fine-tune Arabic transformer encoders across all four subtasks. Memes additionally use a vision tower (ViT-B/16 for the binary subtask, CLIP ViT-L/14 for the fine-grained one) late-fused with MARBERTv2; ensembling choices and decision thresholds are tuned per subtask throughout. Two methodological choices account for much of our gains: selecting checkpoints by development macro-F1 rather than loss, and aligning the B2 training target with the evaluation distribution. Our largest single effect, however, was diagnostic. The B1 blind test set is approximately class-balanced while its training data is ~80/20, so our dev-tuned threshold collapsed macro-F1 from 0.9557 on dev to 0.6345 on test. We recovered the test class balance algebraically from the aggregate metrics the platform reported to all participants; re-thresholding the same saved probabilities, with no retraining, restored 0.8628. Because this uses only aggregate leaderboard feedback (within the task rules), the B1 figure is not directly comparable to teams that did not use the channel and is not a generalisation estimate. This correction is also not general: on the weaker multi-label classifier (A2), stronger prior matching reduced performance in every setting we tested, so matching the test prior is not necessarily F1-optimal. We additionally report a catalogue of quantified negative results, including external-data augmentation with AraSafe. Final test macro-F1: A1 0.7869, A2 0.3463, B1 0.8628, B2 0.7622.</abstract>
<identifier type="citekey">wahid-etal-2026-dynamos</identifier>
<location>
<url>https://aclanthology.org/2026.arabicnlp-sharedtasks.46/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>371</start>
<end>378</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Dynamos at ArGuard Shared Tasks: Multimodal Transformers and Selective Ensembling for Binary and Fine-Grained Arabic Hateful Meme and Harmful Prompt Detection
%A Wahid, Sheikh Abdul
%A Yamin, Muhammad
%A Arshad, Ali
%Y Abdul-Mageed, Muhammad
%Y Elsayed, Tamer
%Y Bouamor, Houda
%Y El-Beltagy, Samhaa
%Y Al-Khalifa, Hend
%Y Abdelali, Ahmed
%Y Elmadany, AbdelRahim
%Y El Mekki, Abdellah
%Y Darwish, Kareem
%Y Alkhereyf, Sakhar
%Y Luqman, Hamzah
%Y Khalifa, Salam
%Y Magdy, Samar
%Y Alyafeai, Zaid
%Y Hamed, Injy
%Y Zaghouani, Wajdi
%Y Zitouni, Imed
%S Proceedings of The Fourth Arabic Natural Language Processing Conference: Shared Tasks
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-497-2
%F wahid-etal-2026-dynamos
%X For the ArGuard 2026 shared task on harmful content detection in Arabic memes (Task A) and LLM prompts (Task B), we fine-tune Arabic transformer encoders across all four subtasks. Memes additionally use a vision tower (ViT-B/16 for the binary subtask, CLIP ViT-L/14 for the fine-grained one) late-fused with MARBERTv2; ensembling choices and decision thresholds are tuned per subtask throughout. Two methodological choices account for much of our gains: selecting checkpoints by development macro-F1 rather than loss, and aligning the B2 training target with the evaluation distribution. Our largest single effect, however, was diagnostic. The B1 blind test set is approximately class-balanced while its training data is ~80/20, so our dev-tuned threshold collapsed macro-F1 from 0.9557 on dev to 0.6345 on test. We recovered the test class balance algebraically from the aggregate metrics the platform reported to all participants; re-thresholding the same saved probabilities, with no retraining, restored 0.8628. Because this uses only aggregate leaderboard feedback (within the task rules), the B1 figure is not directly comparable to teams that did not use the channel and is not a generalisation estimate. This correction is also not general: on the weaker multi-label classifier (A2), stronger prior matching reduced performance in every setting we tested, so matching the test prior is not necessarily F1-optimal. We additionally report a catalogue of quantified negative results, including external-data augmentation with AraSafe. Final test macro-F1: A1 0.7869, A2 0.3463, B1 0.8628, B2 0.7622.
%U https://aclanthology.org/2026.arabicnlp-sharedtasks.46/
%P 371-378
Markdown (Informal)
[Dynamos at ArGuard Shared Tasks: Multimodal Transformers and Selective Ensembling for Binary and Fine-Grained Arabic Hateful Meme and Harmful Prompt Detection](https://aclanthology.org/2026.arabicnlp-sharedtasks.46/) (Wahid et al., ArabicNLP 2026)
ACL