WizardLM: GPT-4-evolved instructions for fine-tuning

To train a model to follow complex instructions, you need many examples of difficult instructions. But collecting this data is slow and expensive. WizardLM solves the problem by asking GPT-4 to "evolve" simple instructions, making them more complex automatically.

The process is called Evol-Instruct: a basic instruction is expanded, made more specific, or given additional constraints, until it becomes much more demanding than the original.

The model fine-tuned on this evolved data reaches performance very close to ChatGPT on evaluation benchmarks, showing that instruction complexity matters as much as quantity.