Learn Before
Analysis of Pointwise Rating Loss Behavior
A reward model is being trained using the objective function L = - (s - r)^2, where s is the target score and r is the model's predicted reward. During an evaluation step, you observe two distinct predictions for two different data points:
- Data Point 1: Target score
s_1 = 7, Model predictionr_1 = 6 - Data Point 2: Target score
s_2 = 4, Model predictionr_2 = 5
Compare the loss values (L_1 and L_2) for these two data points. Based on this comparison, explain what this specific objective function incentivizes the model to do regarding the magnitude versus the direction of its prediction errors.
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
A reward model is being trained using an objective function that aims to maximize the value of for each data point, where 's' is a fixed target score and 'r' is the model's predicted reward. For a particular data point, the target score 's' is 5.0. The model currently predicts a reward 'r' of 4.5. How would the value of L for this data point change if the model's prediction were updated to 5.5?
Analysis of Pointwise Rating Loss Behavior
A machine learning engineer is training a reward model. The goal is for the model's output,
r, to be as close as possible to a set of human-provided target scores,s. The engineer chooses the following objective function to maximize for each data point:L = - (s - r)^2. Why is maximizing this objective function an effective strategy for achieving the engineer's goal?