RoSafetyBench: A Culturally-Native Safety Benchmark for Romanian Language Models
收藏Zenodo2026-05-07 更新2026-05-29 收录
下载链接：
https://zenodo.org/doi/10.5281/zenodo.20070248
下载链接
链接失效反馈官方服务：
资源简介：
RoSafetyBench is a dedicated, culturally-native safety benchmark for Romanian language models, authored as native Romanian rather than translated from English. It comprises 953 prompts across six safety dimensions (toxicity, bias, hallucination, jailbreak resistance, over-refusal, and cross-lingual transfer), designed around Romanian-specific societal risks including anti-Roma discrimination, regional stereotypes, and Romanian profanity patterns.
This deposit contains the v1 benchmark data (6 JSONL files) and a README describing the schema and intended use.
提供机构：
Zenodo创建时间：
2026-05-07
