Tiny Aya Safety Explorer
Saved outputs of Tiny Aya Water before and after layer-restricted vs broad safety LoRA, on 228 Japanese/English validation prompts. Every answer and verdict shown here is a genuine output of the experiment (no live model). Code and write-up.
Content warning: the prompts are jailbreak-style requests about crime, violence, self-harm and similar topics, and a few answers were judged unsafe. Labels come from the automatic judge Qwen3Guard-Gen-0.6B and have not been reviewed by humans. Validation split, not the final heldout test.
Headline metrics
| Metric | Original | Restricted LoRA (layer 18) | Broad LoRA (all layers) |
|---|
