Drag and Drop Question You manage an Azure Al Foundry project. You fine-tune the base model. During evaluation, you observe that the model is overfitting and its responses are highly varying. You…

The correct answer is learning_rate_multiplier; batch_size. To improve a fine-tuned model that is overfitting and exhibiting high variance, adjust the learning_rate_multiplier to reduce overfitting and the batch_size to reduce variance.

Optimize language models for AI applications

Question

Drag and Drop Question You manage an Azure Al Foundry project. You fine-tune the base model. During evaluation, you observe that the model is overfitting and its responses are highly varying. You need to improve the fine-tuned model. Which hyperparameters should you use? To answer, move the appropriate hyperparameters to the correct requirements. You may use each hyperparameter once, more than once, or not at all. You may need to move the split bar between panes or scroll to view content. NOTE: Each correct selection is worth one point. Answer:

Exhibit

Answer Area

Drag items

seedbatch_sizelearning_rate_multiplier

Correct arrangement

learning_rate_multiplier
batch_size

Explanation

To improve a fine-tuned model that is overfitting and exhibiting high variance, adjust the learning_rate_multiplier to reduce overfitting and the batch_size to reduce variance.

Approach. To answer this question, you need to drag the 'learning_rate_multiplier' hyperparameter to the 'Reduce overfitting' requirement box and the 'batch_size' hyperparameter to the 'Reduce variance' requirement box.

Reduce overfitting: Overfitting occurs when a model learns the training data too well, including noise, leading to poor generalization. A lower learning_rate_multiplier (resulting in a smaller learning rate) means smaller updates to the model's weights during training. This can prevent the model from rapidly adjusting to noise in the training data and making drastic changes that lead to over-specialization, thereby helping to mitigate overfitting and improve generalization.
Reduce variance: High variance in model responses suggests instability and sensitivity to small fluctuations in the input data or training process. A larger batch_size during training leads to more stable and less noisy gradient estimates. This stability helps the model converge more smoothly and less erratically, making its predictions less sensitive to individual data points and thus reducing the variance in its responses.

Common mistakes.

common_mistake. Dragging 'seed' to either requirement is incorrect because seed is primarily used for reproducibility of experiments, ensuring that random processes (like weight initialization or data shuffling) yield the same results across multiple runs. It does not directly impact or reduce overfitting or variance in the model's performance. Similarly, swapping learning_rate_multiplier and batch_size would be incorrect, as batch_size is more directly related to the stability of gradient estimates and reducing output variance, while learning_rate_multiplier influences the step size of weight updates, which is crucial for preventing overfitting.

Concept tested. Understanding and applying hyperparameters in machine learning model training, specifically for addressing overfitting and high variance in the context of fine-tuning large language models or similar deep learning models.

Topics

#Hyperparameter tuning#Overfitting#Model fine-tuning#Generative AI

Community Discussion

No community discussion yet for this question.

Full DP-100 Practice