Achieving 10,000x training data reduction with high-fidelity labelsresearch.google154 pointsbadmonstera year ago