跳到正文
arXiv cs.AI· Mohamad Zbib, Mohamad Bazzi, Ammar Mohanna·· 4 小时前AI 评分26

阿拉伯语安全对齐中的选择性拒绝:SFT、DPO 与 Guard 校准的实证研究

Arabic Safety Alignment as Selective Refusal: An Empirical Study of SFT, DPO, and Guard Calibration

AI 导读

针对阿拉伯语大模型的安全对齐研究在 AraSafe 数据集上评测了 5 个模型共 130 次运行,发现仅用拒绝样本做 SFT 会退化为一律拒绝,而精选的混合 SFT 配置可在良性拒绝率 B 为 14%–23% 时把有害提示拒绝率 H 提到 90%–93%。

正文 · 原文

View PDF HTML (experimental)

Abstract:Arabic large language models must refuse harmful prompts without over-refusing benign or sensitive prompts, yet a single refusal rate hides this trade-off. We evaluate it using benign refusal B and harmful-prompt refusal H, where H measures refusal rather than harmful compliance. Across five Arabic-capable models and 130 runs on the full human-written AraSafe set, refusal-only supervised fine-tuning (SFT) collapses toward blanket refusal, whereas selected mixed-SFT configurations reach H = 90% to 93% at B = 14% to 23%; four selected configurations exceed the H = 90% target in all three runs, while Fanar does so in two of three. Direct Preference Optimization (DPO) and inference guards change B and H differently across models rather than acting as uniform upgrades. In a blinded 300-response audit, annotator binary-refusal agreement is 89.0% (kappa = 0.78); Qwen3Guard and Aya Expanse 32B reach 88.7% and 91.0% accuracy, respectively, with no conclusive paired difference. Selected SFT raises H on Arabizi for all five models, but none reaches 90%, showing only partial transfer from Modern Standard Arabic. Overall, the results support model-specific operating-point selection: set a deployment target and retain only interventions that improve it.
Comments: The Fourth Arabic Natural Language Processing Conference(ArabicNLP 2026)
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2608.29378 [cs.CL]
  (or arXiv:2608.29378v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2608.29378

arXiv-issued DOI via DataCite

Submission history

From: Mohammad Zbib [view email]
[v1] Sat, 29 Aug 2026 17:25:01 UTC (2,241 KB)
[v2] Tue, 17 Nov 2026 13:26:30 UTC (2,241 KB)

来源:arXiv cs.AI · arxiv.org