Incorporating Penalty Terms for Controllable Decoding
A simple yet effective technique for enhancing control over text generation is to integrate penalty terms directly into the decoding process. This approach modifies the search objective to guide the model's output according to specific constraints or desired attributes, making the decoding process more controllable.
0
1
Tags
Ch.5 Inference - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
MBR Decoding as an Alternative to MAP Decoding
Incorporating Penalty Terms for Controllable Decoding
Improving Generic Text Generation
A language model is tasked with generating creative story endings. Its current decoding process consistently produces endings that are grammatically perfect and logically sound, but are often predictable and repetitive (e.g., '...and they all lived happily ever after.'). Which of the following statements best analyzes why modifying the search objective could address this issue?
Diagnosing Repetitive Text Generation
Learn After
Decoding Objective with Penalty Term
A language model is being used to generate one-sentence summaries of news articles. The initial outputs are often too long and contain repetitive phrases (e.g., 'The study showed the research indicated that...'). To improve the quality of the summaries, a penalty term is added to the decoding process. Which of the following penalty strategies would be most effective at addressing both of the identified issues?
Evaluating a Penalty Term for Creative Writing
A language model is exhibiting several undesirable behaviors during text generation. Match each problem with the penalty term specifically designed to mitigate it.