If you’ve been tracking Arabic LLM evaluation, you’ve probably noticed a growing tension: the number of benchmarks and leaderboards is expanding rapidly, but are we actually measuring what we think we’re measuring? A new open-source initiative suggests the answer is a firm no. QIMMA, which means ‘summit’ in Arabic, ran a rigorous quality audit on over 52,000 samples from 14 established Arabic benchmarks before evaluating any models. The results are sobering. As many as 40 percent of samples in some widely-used benchmarks were flagged or eliminated for containing systematic quality issues. We’re not talking about edge cases. The problems included factually wrong gold answers, corrupt text, encoding errors, and answers that didn’t comply with the evaluation protocol. One benchmark had gold indices that simply didn’t match any of the answer choices. Another contained stereotype-reinforcing content that flattened diverse Arab communities into monolithic generalizations. To catch these, QIMMA’s pipeline first used two strong LLMs—Qwen3-235B and DeepSeek-V3—to independently score each sample against a 10-point rubric. If either model scored a sample below 7 out of 10, it got a ticket to human review. Native Arabic speakers then made the final call, weighing dialectal nuance and cultural context that automated systems might miss. What’s left is a clean evaluation suite that is 99 percent native Arabic content, spans seven domains from healthcare to software development, and includes the first Arabic leaderboard with code evaluation. The team adapted HumanEval+ and MBPP+ with Arabic problem statements. The implication is uncomfortable but clear: a good chunk of the Arabic NLP leaderboard you’re looking at today is likely contaminated by bad data. QIMMA offers a cleaner starting point, but it also raises a sharp question about every other benchmark we take for granted.