AK_QMUL at ArGuard Shared Tasks: A Single Fine-Tuned JAIS for Arabic Harmful-Prompt Detection

AK_QMUL's single JAIS-6.7B backbone, fine-tuned with 4-bit QLoRA and a linear classification head over its last-token representation, ranked first on both tracks of the ArGuard-2026 shared tasks on Arabic harmful-prompt detection, scoring macro-F1 0.984 on the binary safe/unsafe task (B1) and 0.790 on the seven-way harm-domain task (B2). For B1 the team trained a four-way head factoring labels by dialect register ({safe, unsafe} x {MSA, other}), augmented the training fold with MSA prompts from AraSafe, and decoded with plain argmax; for B2 it averaged softmax outputs of five random seeds. On B1 the model was accurate across all six dialects (93.6–99.8%) with predicted unsafe rates tracking true rates closely, while on B2 six of seven classes reached F1 between 0.73 and 0.98, with residual error concentrated in the Bully class, which overlaps semantically with harassment and harm-to-others.

arXiv:2610.36360v1 Announce Type: new Abstract: We present AK QMUL's submissions to the ArGuard-2026 shared tasks on Arabic harmful-prompt detection: a binary safe/unsafe task B1 and a seven-way harm-domain task B2 . Both submissions use one backbone -- JAIS-6.7B fine-tuned with 4-bit QLoRA and a linear classification head over its last-token representation. For B1 we train a four-way head that factors the label by dialect register {safe, unsafe} x {MSA, other}, we augment the training fold with MSA prompts from AraSafe, and decode with a plain argmax to score safe/unsafe. For B2 we fine-tune the same backbone as a seven-way classifier and average the softmax outputs of five random seeds. Both submissions ranked first on their track, with macro-F1 0.984 on B1 and 0.790 on B2. We report results by dialect and by class: on B1 the model is accurate across all six dialects 93.6--99.8% and its predicted unsafe rate tracks the true rate closely; on B2 six of the seven classes reach F1 between 0.73 and 0.98, with the residual error concentrated in the Bully class, which overlaps semantically with harassment and harm-to-others.