LLM Training & Fine-tuning QUIZ (MCQ QUESTIONS AND ANSWERS)

Counting parameters

Reducing the precision of model weights to decrease size and increase speed

Increasing precision

Deleting weights

It requires more resources

It fine-tunes only a small subset of parameters, reducing computational cost

It deletes most parameters

It makes models larger

Replacing the entire model

Deleting layers

Random layer insertion

Adding small trainable modules (adapters) between frozen layers

Manually writing prompts

Learning continuous prompt embeddings while keeping the model frozen

Deleting prompts

Hardware tuning

Random instruction generation

Fine-tuning models on diverse instruction-following datasets to improve task generalization

Ignoring instructions

Only mathematical instructions

Random learning

Ignoring feedback

Training models using human preferences as rewards to align with human values

Only supervised learning

To give financial rewards

To delete bad outputs

To slow down training

To predict human preferences and guide model training

A data structure

A programming language

A hardware setting

A reinforcement learning algorithm used in RLHF for stable policy updates

Transferring knowledge from a larger model to a smaller one

Water purification

Deleting models

Combining models randomly

A school program

A larger "teacher" model trains a smaller "student" model

Two equal models

Deleting the teacher

A parameter-efficient fine-tuning method that adds trainable low-rank matrices

A new model architecture

A data preprocessing technique

A hardware component

Growing the network

Adding more layers

Removing less important weights or neurons to reduce model size

Random deletion

To delete data

To count parameters

To measure how well the model's predictions match the actual targets

To speed up training

Regression tasks

Classification tasks

Clustering

Sorting

Weights getting old

A regularization technique that penalizes large weights to prevent overfitting

Random weight deletion

Increasing weights

Limiting gradient values to prevent exploding gradients

Deleting gradients

Increasing gradients

Random gradient modification

Gradients becoming too large

Gradients becoming extremely small, hindering learning in deep networks

Gradients disappearing visually

No gradients exist

Gradients becoming too small

Gradients becoming extremely large, causing unstable training

Physical explosions

Gradients staying constant

Training on one task only

Deleting tasks

Sequential single-task training

Training a model on multiple related tasks simultaneously

To gradually increase the learning rate at the beginning to stabilize training

To heat up the hardware

To decrease learning rate immediately

To stop training early

An optimization algorithm that minimizes loss by updating weights in the direction of steepest descent

Increasing model weights randomly

Deleting gradients

A type of neural network

Transfer learning

Random initialization

Data deletion

Hardware upgrading

When the model performs poorly on training data

When the model generalizes well to new data

When the model learns noise in the training data and fails to generalize

When the model runs out of computational power

Increasing the batch size significantly

Using regularization techniques like dropout or early stopping

Training for more epochs

Removing validation data

Training without labels

Fine-tuning with labeled input-output pairs

Unsupervised learning only

Random weight adjustment

To increase training data

To slow down training

To delete bad data

To monitor model performance and prevent overfitting

How fast the model processes data

The speed of the computer

The step size for updating model weights during optimization

The rate of data collection

Training becomes more stable

The model may fail to converge or become unstable

Training speeds up safely

The model becomes more accurate

A single data point

One complete pass through the entire training dataset

A type of model architecture

A hardware component

The total size of the dataset

The model's memory size

The number of samples processed before updating model weights

The output size