Post
79
🤬Multilingual AI models are easier to misuse in low-resource languages.
For the @Lexsi Labs AI Safety and Alignment Hackathon I looked into why this happens in @Cohere 's Tiny Aya Global, and how to fix it without retraining the model.
The short version: the model recognises a harmful request equally well in all ten languages I tested. What weakens outside English is the internal signal that turns that recognition into a refusal. A 33 KB adapter that restores that signal cut harmful answers from 5.4 % to 0.1 % and closed the gap between languages, with no loss in capability or speed.
The full write-up, with the method, results, limitations, an interactive demo and all the open artifacts, is available here:
🤗 https://huggingface.co/blog/davidquicast/xl-refusal-bridge-closing-tiny-aya-cross-lingual-s
For the @Lexsi Labs AI Safety and Alignment Hackathon I looked into why this happens in @Cohere 's Tiny Aya Global, and how to fix it without retraining the model.
The short version: the model recognises a harmful request equally well in all ten languages I tested. What weakens outside English is the internal signal that turns that recognition into a refusal. A 33 KB adapter that restores that signal cut harmful answers from 5.4 % to 0.1 % and closed the gap between languages, with no loss in capability or speed.
The full write-up, with the method, results, limitations, an interactive demo and all the open artifacts, is available here:
🤗 https://huggingface.co/blog/davidquicast/xl-refusal-bridge-closing-tiny-aya-cross-lingual-s