Learn Before
Analyzing Model Error with Plackett-Luce Loss
A model is being trained to rank four items (A, B, C, D) for a given query. The ground-truth preference is A > B > C > D. The model outputs scores that result in the predicted ranking A > B > D > C. The training process aims to minimize the negative log-likelihood of the ground-truth sequence. Analyze which specific step in the ground-truth sequence generation contributes the most to the loss for this training example, and explain why.
0
1
Tags
Ch.4 Alignment - 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
Plackett-Luce Loss Formula
A model is being trained for a listwise ranking task. For one training example, it must rank three items: Item X, Item Y, and Item Z. The correct, ground-truth ranking is X > Y > Z. The training objective is to minimize the negative log-likelihood of observing this ground-truth sequence. Which expression correctly represents the quantity to be minimized for this single training instance, where P(A | S) is the probability of choosing item A from the set of available items S?
Analyzing Model Error with Plackett-Luce Loss
In a listwise ranking task, if the training objective is to minimize the negative log-likelihood of the ground-truth ranked sequences, a decrease in the loss value over training epochs signifies that the model is assigning a lower probability to the correct sequences.