According to @gdb, ML debugging uncovered data leakage and eval flaws, highlighting fixes for training pipelines and reproducible benchmarks.