You want to teach a large language model to identify the capital of a country using a few examples. Construct a complete prompt that provides the capitals for France, Japan, and Germany as demonstrations, and then asks the model for the capital of Italy.

Google

A few-shot learning prompt can utilize a simple pattern to guide a model in a translation task. For instance, to translate Chinese words into English, several demonstrations are provided to establish the input-output format, followed by a new user query for the model to process. An example of such a prompt is:

**DEMO**
现在 → now
来 → come
去 → go
男孩 → boy

**USER**
女孩 →

Example of a Few-Shot Prompt for Chinese-to-English Translation

A user provides the following text to a large language model to teach it a simple rule for making words plural:

cat → cats
dog → dogs
book → books
fish →

Based on the specific pattern established by the examples provided, what is the most probable output the model will generate for the last line?

A user is trying to get a language model to provide the Spanish translation for English animal names. The user is surprised when the model gives an unpredictable response for the last line. Analyze the text provided by the user. Identify the specific line that breaks the established pattern and explain why this inconsistency would likely confuse the model.

Learn Before

Related