analiza aiqbench
This study exposes critical safety disparities in open-weight models when deployed globally, revealing that refusal rates drop significantly for low-resource languages like Somali compared to English benchmarks. The findings highlight a systemic gap where non-refusal outputs often manifest as incoherent or off-topic generations rather than fluent harmful compliance, challenging the reliability of current alignment strategies outside dominant linguistic environments.
SomaliBench Eval: Measuring English-to-Somali Refusal Gaps in Open-Weight Language Models
arXiv:2605.25420v2 Announce Type: replace Abstract: Large language model safety evaluation remains heavily English-centered, leaving low-resource languages under-measured even when models are deployed globally. We evaluate four open-weight instruction-tuned models on SomaliBench v0, a native-author-verified benchmark of 100 harmful-intent prompts paired across English and Somali. Each of Llama-3.1-8B-Instruct, Gemma-2-9B-Instruct, Qwen-2.5-7B-Instruct, and Aya-23-8B is run locally with temperature 0 and the same English "helpful, harmless, and honest" (HHH) system prompt. We find large English-to-Somali refusal gaps for all four models, ranging from 0.40 to 0.93, all strictly positive under a paired bootstrap and significant by exact McNemar tests. For three models, the dominant Somali non-refusal mode is not fluent harmful compliance but unclear output: wrong-language, incoherent, or off-topic generations. A pinned Claude Sonnet snapshot (claude-sonnet-4-5-20250929) classifies each response as refused, complied, or unclear; its safety layer declined 34 of 800 classifications, which the native author labeled manually. A native-author spot-check achieves 100% agreement with the judge (Cohen's $\kappa=1.00$) on 74 comparable rows. We report aggregate refusal rates, category gaps, and reliability statistics only; raw model generations are retained locally and are not released because some may contain harmful content.
Abstract:Large language model safety evaluation remains heavily English-centered, leaving low-resource languages under-measured even when models are deployed globally. We evaluate four open-weight instruction-tuned models on SomaliBench v0, a native-author-verified benchmark of 100 harmful-intent prompts paired across English and Somali. Each of Llama-3.1-8B-Instruct, Gemma-2-9B-Instruct, Qwen-2.5-7B-Instruct, and Aya-23-8B is run locally with temperature 0 and the same English "helpful, harmless, and honest" (HHH) system prompt. We find large English-to-Somali refusal gaps for all four models, ranging from 0.40 to 0.93, all strictly positive under a paired bootstrap and significant by exact McNemar tests. For three models, the dominant Somali non-refusal mode is not fluent harmful compliance but unclear output: wrong-language, incoherent, or off-topic generations. A pinned Claude Sonnet snapshot (claude-sonnet-4-5-20250929) classifies each response as refused, complied, or unclear; its safety layer declined 34 of 800 classifications, which the native author labeled manually. A native-author spot-check achieves 100% agreement with the judge (Cohen's $\kappa=1.00$) on 74 comparable rows. We report aggregate refusal rates, category gaps, and reliability statistics only; raw model generations are retained locally and are not released because some may contain harmful content.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
arXiv:2605.25420 [cs.CL]
(or arXiv:2605.25420v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2605.25420 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 282 cuvinte
