Few-Shot GPT-4 Turbo: Better Results

Tip: I’ve been experimenting with GPT-4 Turbo (128k context window) and found that providing just 3-5 carefully crafted few-shot examples often yields better results than lengthy, detailed instructions for complex data transformations. For instance, when trying to extract specific fields from a messy CSV file, using 5 examples of correctly formatted output dramatically improved the model’s accuracy to 85% compared to a detailed paragraph of instructions that only achieved 68%. It seems the model learns the *pattern* faster from the examples rather than being explicitly told what to do.