QIMMA قِمّة ⛰: A Quality-First Arabic LLM Leaderboard

TII's QIMMA leaderboard applies a multi-stage quality validation pipeline to Arabic benchmarks before evaluating models, discarding samples with systematic issues such as false gold answers and cultural misalignment.

It consolidates 109 subsets from 14 benchmarks into over 52,000 samples across seven domains, including the first Arabic leaderboard with code evaluation.

The post reports discard rates per benchmark, with ArabicMMLU losing 3.1% of its samples.