Benchmark tasks based on real code changes from actual production repositories, not synthetic examples.