Model Benchmarks

One deterministic 500-task evaluation: arithmetic, numeric reasoning, knowledge, coding, data manipulation, language and instruction following: run against every Fikra model. A run publishes exactly when it finishes 500/500.