Orca 2

Orca 2 is a small language model from Microsoft Research that achieves performance far exceeding its size through careful reasoning capability training. Using a technique called explanation tuning where the model learns different reasoning strategies for different tasks, Orca 2 13B outperforms much larger models on complex reasoning benchmarks. Microsoft Research designed Orca 2 to demonstrate that smaller models can match large model reasoning when trained appropriately. Published with open weights for research use.