Learn Before
Origin of Zero-Shot Learning Ability in LLMs
The capacity for zero-shot learning in Large Language Models is not an explicitly programmed feature but rather an emergent property that develops during the pre-training and/or fine-tuning phases.
0
1
Tags
Ch.2 Generative Models - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Iterative Prompt Adjustment in Zero-Shot Learning
Example of a Persona-based Prompt for Grammar Correction
Origin of Zero-Shot Learning Ability in LLMs
Example of a Zero-Shot Prompt for Grammar Correction
A developer wants a large language model to classify customer feedback. They provide the model with the following prompt:
You are an expert sentiment analysis system. Classify the following customer review as 'Positive', 'Negative', or 'Neutral'. Provide only the label. Review: 'The battery life is impressive, but the screen is too dim.'Which of the following statements best explains why this approach tests the model's ability to generalize to a new task based on instructions alone?Revising a Prompt for Generalization
A research team is testing a large language model's ability to perform a task it has not been specifically trained on: summarizing news articles into a single sentence. Which of the following prompts is a clear example of a zero-shot approach?
Learn After
A team of engineers trains a new large-scale language model on a massive and diverse dataset of text from the internet. After training, they are surprised to find that the model can accurately translate sentences from English to French, even though it was never explicitly given English-to-French translation examples. Which statement best analyzes the origin of this unexpected capability?
Evaluating Claims about LLM Capabilities
A large language model's ability to perform tasks it has never been specifically trained on is primarily achieved by adding a specialized 'zero-shot capability module' after its initial pre-training is complete.