amalia-llm/amaliaguard-4b
AMALIAGuard-4B is a 4 billion parameter content safety guard model developed by amalia-llm, fine-tuned from Qwen/Qwen3Guard-Gen-4B with a 32768 token context length. It specializes in classifying user prompts and assistant responses for safety in European Portuguese (pt-PT), uniquely incorporating six GDPR-specific risk categories alongside universal harm categories. This model is optimized for identifying potential regulatory risks and universal safety concerns in LLM pipelines, particularly for applications requiring GDPR compliance in Portuguese-speaking contexts.
Loading preview...
AMALIAGuard-4B: European Portuguese Content Safety Guard
AMALIAGuard-4B is a 4 billion parameter content safety guard model, fine-tuned from Qwen/Qwen3Guard-Gen-4B, specifically designed for European Portuguese (pt-PT). It addresses a critical gap in existing guard models by integrating six GDPR-specific risk categories alongside standard universal harm categories, making it highly relevant for applications requiring compliance with European data protection regulations.
Key Capabilities & Features
- Dual-Pillar Safety Taxonomy: Classifies content across 12 categories, including:
- G-Pillar (GDPR Compliance): Covers Personal Data Extraction, Special Category Data, Consent Violation, Data Subject Rights Obstruction, Unlawful Cross-Border Transfer, and Automated Profiling.
- U-Pillar (Universal Safety): Includes Sexually Explicit Content, Hate Speech, Dangerous Content, Harassment, Violence, and Obscenity/Profanity.
- Language Specialization: Optimized for pt-PT, with U-pillar categories adapted for European Portuguese cultural and linguistic context.
- Operator-Configurable Categories: Supports dynamic category evaluation, allowing operators to limit assessment to specific categories relevant to their deployment by modifying the prompt.
- Robust Training: Fine-tuned on a three-layer synthetic AART pipeline in pt-PT and English, augmented with translated subsets of WildGuardMix and ToxicChat for broader generalization.
Performance & Limitations
- Achieves 99.65% overall F1 on in-domain pt-PT held-out test sets, significantly outperforming zero-shot baselines.
- GDPR violation recall remains an open problem, reflecting the complexity of legal reasoning required.
- Limitations: G-pillar flags are indicators of potential risk, not legal determinations. There is a synthetic-to-real gap, and coverage for other EU languages is untested.
Ideal Use Cases
- LLM pipelines requiring content moderation in European Portuguese.
- Applications needing to identify potential GDPR compliance risks in user interactions or generated content.
- Content filtering for platforms targeting Portuguese-speaking users, especially within the EU regulatory framework.