SFT datasets are lists of these pairs, sometimes with system prompts or multi-turn history. Quality of the target response matters more than volume of noisy pairs. Synthetic pairs from a stronger model are common when human writing does not scale.
One instruction-tuning example: what to do, and the desired output.
SFT datasets are lists of these pairs, sometimes with system prompts or multi-turn history. Quality of the target response matters more than volume of noisy pairs. Synthetic pairs from a stronger model are common when human writing does not scale.