Every assistant claims a benchmark. Most benchmarks measure the benchmark. This is the evaluation set we use internally and why we publish the method rather…
