Reproducing Self-Adapting Language Models