MiJaBench: Bilingual Adversarial Prompts for LLM Safety Auditing

MiJaBench is a bilingual English-Portuguese adversarial benchmark introduced in a paper accepted to Findings of ACL 2026. The dataset is designed for auditing demographic-specific vulnerabilities in large language model safety alignment. It was created by AKCIT and last updated on Hugging Face in June 2026.

Use Cases

• Auditing demographic-specific safety vulnerabilities in LLMs based on the adversarial prompts.
• Benchmarking the robustness of safety alignment techniques across different languages.
• Researching the selective safety trap phenomenon described in the associated paper.

Strengths

• Dataset is associated with peer-reviewed research accepted to Findings of ACL 2026.
• Focuses on a specific, documented safety issue: the selective safety trap in LLM alignment.
• Contains bilingual (English-Portuguese) adversarial prompts, enabling cross-lingual analysis.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.

Provenance

Source

AKCIT, associated with the GitHub repository github.com/iagoalvesb/mijabench.

Collection Method

Likely contains curated adversarial prompts designed for safety testing.

Freshness

Last updated 2026-06-03 23:51:51; freshness should be verified.