Hugging Face fine-tuning in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Model Choice

intermediate

Choosing the right model for fine-tuning

You want to fine-tune a Hugging Face transformer model for a text classification task with limited labeled data. Which model is best suited to start with?

A A pretrained BERT model with a classification head

B A pretrained word2vec embedding model without a transformer

C A convolutional neural network trained on images

D A large GPT-style causal language model pretrained on general text

💡 Hint

Think about models designed for classification tasks and pretrained on language data.

BERT models are pretrained with masked language modeling and are commonly fine-tuned for classification tasks by adding a classification head. GPT models are mainly for generation, CNNs are for images, and word2vec lacks transformer architecture.

❓ Predict Output

intermediate

Output of training metrics during fine-tuning

Consider this snippet for fine-tuning a Hugging Face model using Trainer API:

from transformers import Trainer, TrainingArguments

training_args = TrainingArguments(output_dir='./results', num_train_epochs=1, per_device_train_batch_size=2)
trainer = Trainer(model=model, args=training_args, train_dataset=train_dataset)
train_result = trainer.train()
print(train_result.metrics)

What is the expected type and content of train_result.metrics after training?

A A list of loss values for each training step

B A string summarizing the training progress

C A dictionary containing keys like 'train_loss' and 'train_runtime' with float values

D An integer representing the total number of training steps

💡 Hint

Think about what the Trainer API returns after training completes.

The Trainer's train() method returns a TrainOutput object whose 'metrics' attribute is a dictionary with keys like 'train_loss', 'train_runtime', and 'train_samples_per_second'.

❓ Hyperparameter

advanced

Choosing learning rate for fine-tuning

You are fine-tuning a pretrained transformer model on a small dataset. Which learning rate is most appropriate to avoid overfitting and unstable training?

A 1.0 (extremely high learning rate)

B 5e-5 (a small learning rate commonly used for fine-tuning)

C 0.1 (very high learning rate)

D 1e-1 (which equals 0.1, very high)

💡 Hint

Fine-tuning usually requires smaller learning rates than training from scratch.

Learning rates like 5e-5 are typical for fine-tuning transformers to ensure stable updates and prevent overfitting. Very high learning rates cause training to diverge.

🔧 Debug

advanced

Debugging a mismatch error during fine-tuning

You try to fine-tune a Hugging Face model but get this error:

RuntimeError: The size of tensor a (10) must match the size of tensor b (5) at non-singleton dimension 1

What is the most likely cause?

A The model's output layer size does not match the number of classes in your dataset

B The batch size is too large for your GPU memory

C The learning rate is set too high causing unstable gradients

D The tokenizer vocabulary size is smaller than the model's embedding size

💡 Hint

Think about tensor size mismatches related to classification output dimensions.

This error usually happens when the model's final classification layer expects a different number of classes than the dataset provides. For example, model output size 10 vs dataset labels 5.

🧠 Conceptual

expert

Understanding parameter freezing in fine-tuning

When fine-tuning a large pretrained transformer, freezing some layers can help training. Which statement best explains why freezing layers is useful?

A Freezing layers increases the model's capacity to learn new features faster

B Freezing layers automatically increases the learning rate for unfrozen layers

C Freezing layers allows the model to randomly reset weights during training

D Freezing layers reduces the number of trainable parameters, lowering memory use and preventing overfitting on small datasets

💡 Hint

Consider the effect of freezing on training stability and resource use.

Freezing layers means their weights do not update during training, which reduces memory and computation needs and helps avoid overfitting when data is limited.