Continual pre-training checkpoints using ICE for DialLM across Gemma, Llama, and Qwen base models.