d3-edge (0.6B) fine-tuned outperforms d3 (27B) decision model

Took vLLM-SR d3-edge (0.6B), fine-tuned it with simple SFT + LoRA, and compared it against the much larger d3 (27B). With just a few thousand training examples, the 0.6B model actually outperformed the 27B model on...