Questions

Select the correct answer for each question.

1. What is the primary goal of instruction tuning?

A) Reduce model size

B) Improve GPU efficiency

C) Teach a model to follow structured instructions reliably

D) Increase token vocabulary

2. In an instruction dataset, which field is optional but often useful?

A) Instruction

B) Output

C) Input

D) Model name

3. Why is dataset consistency important in supervised fine-tuning?

A) It reduces GPU memory usage

B) It prevents the model from learning ambiguous formatting patterns

C) It increases model parameter count

D) It removes the need for evaluation

4. What is the main purpose of deduplication?

A) Increase training speed by adding more examples

B) Remove repeated samples that could cause overfitting

C) Improve tokenizer performance

D) Compress the dataset file

5. What problem does gradient accumulation solve?

A) It reduces model size

B) It increases token vocabulary

C) It simulates larger batch sizes without increasing memory usage

D) It removes the need for multiple GPUs

6. Why must you monitor token length during preprocessing?

A) Because longer sequences improve creativity

B) Because models cannot process text longer than one sentence

C) Because exceeding the maximum sequence length can cause memory issues

D) Because tokenization changes model architecture

7. If training loss decreases but behavior does not improve meaningfully, what is the most likely issue?

A) The GPU is overheating

B) The dataset lacks diversity or meaningful signal

C) The tokenizer is incorrect

D) The model is too small

8. What is a common cause of Out-of-Memory (OOM) errors?

A) Low learning rate

B) Too small a dataset

C) Excessive sequence length or batch size

D) Using JSONL format

9. What is “catastrophic forgetting”?

A) When the model forgets its tokenizer

B) When the model performs worse on general tasks after fine-tuning

C) When the GPU runs out of memory

D) When the optimizer resets gradients

10. Why is it important to compare base vs fine-tuned outputs?

A) To verify that hardware is functioning

B) To confirm the tokenizer is installed

C) To measure actual behavioral changes

D) To reduce training time

11. Which preprocessing step helps prevent training noise from repeated examples?

A) Tokenization

B) Deduplication

C) Augmentation

D) Mixed precision

12. What is one risk of synthetic dataset generation?

A) It always reduces accuracy

B) It can introduce errors or bias if not filtered

C) It increases GPU memory usage

D) It prevents gradient accumulation

13. What does mixed precision (FP16/BF16) primarily improve?

A) Vocabulary diversity

B) Memory efficiency and training speed

C) Instruction formatting

D) Dataset augmentation

14. If a model starts repeating phrases during inference, what is a likely contributing factor?

A) Too much evaluation data

B) Highly repetitive training outputs

C) Too many GPUs

D) Small tokenizer vocabulary

15. Which statement best describes the relationship between dataset quality and fine-tuning results?

A) Hyperparameters matter more than data

B) Hardware matters more than data

C) Data quality strongly determines behavior quality

D) Dataset size is irrelevant