A boy and his dog.

I can't speak for the development of frontier models, but at 50M parameters data quality is everything . Scout finished her training on TinyStories to 40,000 mark. Prompt probes at checkpoints along the way gave me a window into what she was learning, and integrating Claude Code was surprisingly helpful. Using a language model with a recurring task set to run a prompt probe-based quality check every 500 or so training steps gave me a quantitative metric on a qualitative value. The qualitative me