Zero-shot
The ability of a model to immediately perform a new task using only pre-trained knowledge and natural language instructions (prompts), without separate training data or examples (few-shot).
Detailed explanation
Why it matters in tool selection
Zero-shot performance represents the baseline strength of the AI model. Selecting a model with excellent zero-shot performance reduces the hassle of constructing examples, minimizes prompt length to save API costs (tokens), and shortens latency in real-time services.
What to check
- Does it accurately follow the desired output format (JSON, Markdown, etc.) based only on instructions?
- Does the model sufficiently understand the common sense or basic concepts of the domain?
- Is a Chain-of-Thought technique needed instead of Zero-shot when complex reasoning is required?
Example
Input: "Analyze the sentiment of the following sentence: 'The weather is so nice today, it's perfect for a walk.'" Output: "Positive" (Note: No example sentences for the 'Positive' outcome were provided)
Confusing terms
Few-shot
A method that includes about 2 to 5 input-output examples within the prompt to assist in performing the task.
Fine-tuning
The process of optimizing performance by updating the model's own weights with a specific dataset, rather than modifying the prompt.