Diagnosing Fine-Tuning Failure for Long Contexts
A development team is tasked with adapting a pre-trained language model for a new application. The model was originally trained on text segments with a maximum length of 1,024 tokens. The new application requires the model to process and maintain context over dialogues that are often 3,000 tokens long. To prepare the model, the team fine-tunes it using a large, high-quality dataset where each example is a short, self-contained text segment, averaging 500 tokens. Upon deployment, the team observes that the model performs poorly on the longer dialogues, frequently losing track of earlier parts of the conversation. Based on this information, analyze the team's methodology and identify the primary flaw in their fine-tuning strategy. What specific characteristic should their fine-tuning data have possessed to better achieve their goal?
0
1
Tags
Ch.3 Prompting - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Analysis in Bloom's Taxonomy
Cognitive Psychology
Psychology
Social Science
Empirical Science
Science
Related
A research team has a language model that was pre-trained exclusively on text segments with a maximum length of 2,048 tokens. The team's goal is to adapt this model to accurately summarize legal documents that are frequently 5,000 tokens long, a task at which the model currently performs poorly. Given this specific goal, which of the following fine-tuning strategies is most likely to be effective?
Diagnosing Fine-Tuning Failure for Long Contexts
Designing a Fine-Tuning Strategy for Long-Context Tasks