--- license: mit tags: [weightpatch, weight-editing, lexsi] --- # Tiny Aya Global as released: baseline measurements (WeightPatch) What the released `CohereLabs/tiny-aya-global` does before any edit, measured with the WeightPatch grader's suites and judges. `profile.json`: harmful answers (HarmBench, held-out and perturbed prompts; Granite Guardian and Llama Guard), refusals of safe requests (XSTest, OR-Bench hard, JailbreakBench benign), and accuracy (MMLU, Hindi MMLU, MedMCQA), 100 items each with Wilson intervals. `metric_gaming_baseline.json` and `metric_gaming_judge_labels.jsonl`: School of Reward Hacks prompts, the model's answers and the judge's label for each (gamed, honest, broken).