Embedding dimensionality considerations in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why choose higher embedding dimensions?

When creating embeddings for text data, why might you choose a higher dimensionality?

A To capture more detailed features and subtle differences between items

B To reduce the training time and computational cost

C Because higher dimensions always guarantee better accuracy regardless of data

D To make the embeddings easier to visualize in 2D or 3D plots

💡 Hint

Think about what more dimensions allow the model to represent.

Higher embedding dimensions allow the model to capture more detailed and subtle features of the data, improving representation quality. However, it may increase computational cost and risk overfitting.

❓ Metrics

intermediate

Effect of embedding size on cosine similarity

You have two sets of embeddings: one with 50 dimensions and one with 300 dimensions. Both sets are normalized. Which statement about cosine similarity between embeddings is true?

A Higher dimensional embeddings tend to have cosine similarities closer to zero due to sparsity

B Cosine similarity values are directly comparable regardless of embedding size

C Lower dimensional embeddings always produce higher cosine similarity values

D Cosine similarity cannot be computed for embeddings with different dimensions

💡 Hint

Consider how dimensionality affects vector distribution and sparsity.

Higher dimensional embeddings often become sparser, causing cosine similarity values to cluster closer to zero, making distinctions less pronounced.

❓ Predict Output

advanced

Output shape of embedding layer

Consider the following PyTorch code snippet creating an embedding layer and passing input indices:

Prompt Engineering / GenAI

import torch
import torch.nn as nn

embedding = nn.Embedding(num_embeddings=1000, embedding_dim=128)
input_indices = torch.tensor([[1, 2, 3], [4, 5, 6]])
output = embedding(input_indices)
print(output.shape)

A torch.Size([128, 3, 2])

B torch.Size([3, 2, 128])

C torch.Size([2, 128])

D torch.Size([2, 3, 128])

💡 Hint

Think about how embedding layers map indices to vectors.

The input tensor has shape (2, 3). Each index is replaced by a 128-dimensional vector, so output shape is (2, 3, 128).

❓ Hyperparameter

advanced

Choosing embedding dimension for a small dataset

You have a small dataset with only 500 unique words. Which embedding dimension is most appropriate to avoid overfitting while maintaining useful representation?

A 1024 dimensions to capture all nuances

B 3000 dimensions to ensure maximum expressiveness

C 50 dimensions to balance capacity and overfitting risk

D 5 dimensions because the dataset is small

💡 Hint

Think about the trade-off between model complexity and data size.

For small datasets, moderate embedding sizes like 50 help avoid overfitting while still capturing meaningful features.

🔧 Debug

expert

Why does increasing embedding dimension cause training to fail?

After increasing embedding dimension from 100 to 10000 in a neural network, training loss becomes NaN immediately. What is the most likely cause?

A Higher dimension embeddings always cause NaN due to numerical overflow

B The model runs out of memory causing unstable gradients

C The optimizer does not support large embedding sizes

D Embedding dimension does not affect training stability

💡 Hint

Consider hardware limits and how large parameters affect training.

Very large embeddings increase memory use and can cause unstable gradients or overflow, leading to NaN losses.