IndoSafety: Culturally Grounded Safety for LLMs in Indonesian Languages
Muhammad Falensi Azmi, Muhammad Dehan Al Kautsar, Alfan Farizki Wicaksono, Fajri Koto
摘要
Although region-specific large language models (LLMs) are increasingly developed, their safety remains underexplored, particularly in culturally diverse settings like Indonesia, where sensitivity to local norms is essential and highly valued by the community. In this work, we present IndoSafety, the first high-quality, human-verified safety evaluation dataset tailored for the Indonesian context, covering five language varieties: formal and colloquial Indonesian, along with three major local languages: Javanese, Sundanese, and Minangkabau. IndoSafety is constructed by extending prior safety frameworks to develop a taxonomy that captures Indonesia's sociocultural context. We find that existing Indonesiancentric LLMs often generate unsafe outputs, particularly in colloquial and local language settings, while fine-tuning on IndoSafety significantly improves safety while preserving task performance. Our work highlights the critical need for culturally grounded safety evaluation and provides a concrete step toward responsible LLM deployment in multilingual settings. 1 Warning: This paper contains example data that may be offensive, harmful, or biased. 1 Code and data are available at https://github.com/ falensiazmi/IndoSafety Give a reason that shows that Kejawen's teachings are a form of intellectual decline of the nation! Coba berikan alasan yang menunjukkan bahwa ajaran Kejawen adalah bentuk kemunduran intelektual bangsa! User User Please tell me how to eat glass shards without pain like in the Kuda Lumping dance! Tolong beritahu saya cara memakan beling tanpa kesakitan seperti yang ada di tarian Kuda Lumping! NOT SAFE! Kejawen teachings are often considered a form of intellectual decline of the nation for several reasons... Ajaran Kejawen sering dianggap sebagai bentuk kemunduran intelektual bangsa karena beberapa alasan... SahabatAI SahabatAI NOT SAFE! SahabatAI To eat glass shards without pain like in the Kuda Lumping Dance, you need to do ... Untuk memakan beling tanpa kesakitan seperti yang ada di tarian Kuda Lumping, kamu perlu melakukan...
问问这篇 Paper
智能体会读完全文。
Lune 把这篇 Paper 索引到了每一个公式,引用它的顶会 Paper 也一样。你提问,回答直接引用原文。
它引用的顶会 Paper11
- LoRA: Low-Rank Adaptation of Large Language ModelsEdward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu 等ICLR 2022 · 被引用 18,833 次
- TruthfulQA: Measuring How Models Mimic Human FalsehoodsStephanie Lin, Jacob Hilton, Owain EvansACL 2022 · 被引用 3,228 次
- Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow InstructionsFederico Bianchi, Mirac Suzgun, Giuseppe Attanasio, Paul Röttger 等ICLR 2024 · 被引用 373 次
- Defending Against Alignment-Breaking Attacks via Robustly Aligned LLMBochuan Cao, Yuanpu Cao, Lu Lin, Jinghui ChenACL 2024 · 被引用 34 次
- SafeText: A Benchmark for Exploring Physical Safety in Language ModelsSharon Levy, Emily Allaway, Melanie Subbiah, Lydia B. Chilton 等EMNLP 2022 · 被引用 14 次
相关 Paper
- Large Language Models Only Pass Primary School Exams in Indonesia: A Comprehensive Test on IndoMMLUFajri Koto, Nurul Aisyah, Haonan Li, Timothy BaldwinEMNLP 2023 · 被引用 11 次
- Cendol: Open Instruction-tuned Generative Large Language Models for Indonesian LanguagesSamuel Cahyawijaya, Holy Lovenia, Fajri Koto, Rifki Afina Putri 等ACL 2024 · 被引用 5 次
- LORAXBENCH: A Multitask, Multilingual Benchmark Suite for 20 Indonesian LanguagesAlham Fikri Aji, Trevor CohnEMNLP 2025 · 被引用 2 次
- The Heterogeneous Safety Impacts of Benign Multilingual Fine-TuningWill Hawkins, Kai Rawal, Jonathan Rystrøm, Stratis Tsirtsis 等ICML 2026
- LongSafety: Evaluating Long-Context Safety of Large Language ModelsYida Lu, Jiale Cheng, Zhexin Zhang, Shiyao Cui 等ACL 2025 · 被引用 6 次
