Increasing our dataset mixes has provided even stronger outcomes - with our top two models surpassing DeepSeek-V2-Lite, Google’s Gemma 7B, and Mistral 7B. Here’s how.