kevinamiri/Instructgpt-prompts

Empirical research shows that model outputs vary dramatically—up to 76 percentage points—based on prompt phrasing and position sensitivity (primacy/recency bias), particularly in base and SFT-only models. The collection provides structured templates and verb-based directives (e.g., "Classify," "Generate," "Summarize") designed to exploit token weighting patterns and guide instruction-following models toward consistent performance. Targets InstructGPT and GPT-3.5-instruct models through a three-part prompt architecture: background context, input specification, and explicit action instruction.

533 stars.