CURUPIRA: Clever guard for harm and linguistic prompt mitigation in Brazilian Portuguese

2026-04-12 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Advanced, quick

Summary

Curupira is a Brazilian Portuguese-language guard model developed to mitigate harmful prompt exploitation in Large Language Models, addressing challenges in multilingual safety. This model was created using a three-step methodology involving adaptation, data generation, and fine-tuning. Researchers evaluated Curupira against two open guardrail architectures, demonstrating that targeted fine-tuning significantly improves safety classification for Portuguese prompts. The evaluation also revealed favorable efficiency-performance trade-offs for compact models and minimal degradation during cross-lingual assessment. This work, presented at the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) in Salvador, Brazil, highlights the importance of language-specific safety measures for LLMs.

Key takeaway

For research scientists developing LLMs for diverse linguistic contexts, you should consider implementing language-specific guard models like Curupira. This approach, particularly with targeted fine-tuning, can significantly enhance safety classification for underrepresented languages such as Brazilian Portuguese, ensuring more robust and culturally appropriate model deployment. Prioritize evaluating both efficiency and cross-lingual performance to maintain broad applicability.

Key insights

Targeted fine-tuning improves LLM safety classification for underrepresented languages like Brazilian Portuguese.

Principles

• Multilingual LLM safety is challenging.
• Language-specific guard models enhance safety.

Method

The methodology involves adaptation, data generation, and fine-tuning to create language-specific guard models for prompt mitigation.

In practice

• Develop language-specific guard models.
• Utilize targeted fine-tuning for safety.
• Evaluate cross-lingually for robustness.

Topics

• CURUPIRA
• Large Language Models
• Brazilian Portuguese
• Prompt Mitigation
• Safety Classification

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Security Engineer

Related on AIssential

• Linguistic Symptoms: Augmented Generation by Retrieval and Reasoning in LLMs under Portuguese-English Variation in Medical Contexts — Linguistic variation significantly impacts smaller LLMs in medical contexts, with larger models showing greater cross-language robustness.
• Lispector: Fine-tuning of Language Models for Grammar and Spelling Correction in Brazilian Portuguese — Supervised fine-tuning of smaller models outperforms few-shot prompting for grammar correction in Brazilian Portuguese.
• Beyond English and Evasion: A Human-Annotated Multi-Domain Benchmark for High-Stakes LLM Safety Evaluation in Chinese — Chinese LLM safety requires culturally specific benchmarks to counter linguistic and cultural evasion techniques.
• From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model — LLM safety analysis benefits from paired prompt-to-response severity transitions and joint relevance assessment.
• Multilingual jailbreaking of LLMs using low-resource languages — Multi-turn conversations in low-resource languages can jailbreak LLMs, with translation quality being a key factor.
• Can AI Save India’s Dying Languages? | Building a Tulu LLM Without Training Data Ft. Prathamesh — Structured prompting and custom romanization enable LLMs to generate low-resource languages without extensive training data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.