NEW
Font size
WorksheetsPage 1
Total questions: 10
Worksheet time: 5mins
Which model family is described as encoder-only in modern NLP architectures?
Autoregressive decoders like GPT-4
Encoder–decoder hybrids like T5
Diffusion-based latent models
Transformer encoder models like BERT
Which statement best characterizes a strictly autoregressive decoder during generation?
It computes all output tokens in parallel
It encodes the full input before any decoding begins
It simultaneously attends to future and past tokens
It only sees previously generated tokens sequentially
Top-p (nucleus) sampling primarily aims to achieve which outcome in text generation?
Ensure identical outputs across repeated runs
Guarantee the most likely token every time
Balance creativity with coherence in outputs
Reduce vocabulary size during tokenization
Which decoding method is most likely to produce deterministic and focused outputs?
Beam search with high diversity penalty
Top-p sampling with high nucleus
Random sampling with high temperature
Greedy decoding with low temperature
Which penalty discourages the model from repeating words it has already used frequently in the session?
Top-k truncation of candidate set
Frequency penalty applied during decoding
Presence penalty applied only to new tokens
Temperature scaling before softmax
Tokenization in large language models is best described as which process?
Sampling words from a probability distribution
Converting text into discrete subword tokens
Compressing sentences into fixed-length vectors
Parsing syntax using a context-free grammar
Which configuration most increases randomness and diversity in generated text?
High temperature with random sampling
Low temperature with greedy decoding
Top-p value near zero with sampling
Beam width of one with repetition penalty
Which pair correctly matches popular proprietary model families with their category?
BERT and T5 as proprietary models
GPT-4o and Gemini as proprietary models
Claude and LLaMA as proprietary models
LLaMA and Mistral as proprietary models
External tools like search and code interpreters are typically used in conjunction with LLMs to achieve which benefit?
Augment capabilities for tasks requiring tools
Replace the model’s internal reasoning entirely
Reduce token usage during tokenization steps
Increase training data diversity automatically
Setting temperature to a very low value (e.g., 0.1) during decoding most likely has what effect?
Outputs become more deterministic and focused
Outputs explore more diverse continuations
Vocabulary is expanded during inference
Model ignores previously generated tokens
