Data and checkpoints for Data-Efficient Autoregressive-to-Diffusion Language Models via On-Policy Distillation