Quizzes

Neural Nets & Transformers quiz

24 questions from the Neural Nets & Transformers track. Score 80% or higher to unlock a certificate.

1. Tokens

   1. Why do modern language models use subword tokens instead of whole words?

   • Words cannot be stored on disk
   • A word table cannot handle new or rare spellings; subwords reuse pieces and keep the table finite
   • Characters are illegal in UTF-8
   • Token ids only work for English

2. Byte Pair Encoding

   2. What does one BPE merge do?

   • Deletes rare words from English
   • Glues the most common neighboring pair into a new token
   • Trains the attention heads
   • Converts ids into logits

3. Special Tokens

   3. What is a chat template for?

   • Speeding up the hardware
   • Wrapping roles and turns in the special tokens the model was trained to expect
   • Deleting the system prompt
   • Training BPE from scratch

4. The Embedding Table

   4. What does an embedding table do?

   • Compresses the hardware
   • Maps a token id to a vector by returning that id's row
   • Deletes rare words
   • Computes softmax

5. Positions

   5. Why do transformers need a position signal?

   • Softmax cannot run without it
   • Without it, order is lost and the same words in a new order look the same
   • It replaces the embedding table
   • It trains BPE

6. Attention

   6. In causal self-attention, why are some scores set to a huge negative number?

   • To speed up softmax
   • So token i cannot attend to future tokens j > i
   • Because cosine is negative
   • To implement dropout

7. Multi-Head Attention

   7. What problem do multiple attention heads address?

   • They reduce the vocabulary
   • One softmax mix is a bottleneck; several subspaces can specialize in different relations
   • They delete positional encoding
   • They make tokens into characters

8. The Transformer Block

   8. What do residual connections do in a transformer block?

   • They delete attention
   • They add the block's output back to its input so information (and slopes) can flow through depth
   • They replace softmax
   • They tokenize bytes

9. Residuals and LayerNorm

   9. Why add the layer output back to the input (a residual)?

   • To delete the embedding table
   • So each layer can learn a small change, and slopes still reach the early tokens
   • So softmax becomes greedy
   • To train BPE faster

10. Encoder vs Decoder

    10. Why must a GPT-style decoder hide future tokens?

    • Hardware cannot store them
    • Otherwise next-token training would see the answer and the game would be fake
    • Softmax requires a triangle
    • Encoders cannot run without a GPU

11. Pretraining

    11. What is the standard pretraining task for GPT-style models?

    • Image classification of tokens
    • Predict the next token given the previous tokens
    • Human labels on every sentence
    • Minimizing cosine on a knowledge graph

12. Logits and Unembedding

    12. What is a logit in a language model?

    • A guaranteed probability
    • A raw score for one vocab id, before softmax
    • A BPE merge rule
    • A residual connection

13. Decoding

    13. When should an agent decode tool names greedily?

    • Never; always T=2 for creativity
    • When a single valid action is required and randomness is a bug
    • Only if the vocab has four words
    • Greedy is impossible with softmax

14. The KV Cache

    14. What does the KV cache store?

    • The tokenizer merge list
    • Keys and values for tokens already processed, so new tokens need not recompute them
    • The test set
    • Softmax temperature

15. Context Windows

    15. What is a common dangerous truncation bug?

    • Deleting the latest tool result
    • Dropping the oldest messages and losing the pinned system spec
    • Using greedy decoding
    • Counting characters instead of GPUs

16. Prefill and Decode

    16. Why is time-to-first-token mostly prefill?

    • Decode cannot start on GPUs
    • The model must read the whole prompt and fill the cache before it can score the first new token
    • Softmax runs only once ever
    • BPE happens after the last token

17. Fine-Tuning

    17. What does SFT mainly copy?

    • The KV cache
    • The demonstrations you showed: format, style, and any mistakes in them
    • The retrieval index
    • Temperature

18. LoRA

    18. What is LoRA for?

    • Replacing the tokenizer
    • Learning a small low-rank patch instead of updating the full weight matrices
    • Increasing temperature
    • Clustering embeddings

19. Limitations

    19. Why do transformers hallucinate facts?

    • Softmax is random by law
    • Pretraining optimizes plausible next tokens, not grounded truth
    • Attention is always bidirectional
    • LoRA deletes the embedding table

20. Lost in the Middle

    20. Where should the latest tool observation usually sit in a long prompt?

    • In the middle, to be fair to all tokens
    • Near the end (and keep the spec pinned at the front)
    • Only inside the embedding table
    • After the EOS token

21. Scaling

    21. What do scaling laws mainly predict?

    • That bigger models always follow your company policy
    • That loss tends to fall smoothly as you scale params, data, and compute together
    • That attention becomes linear
    • That retrieval is obsolete

22. Padding and Masks

    22. What is an attention mask for padding doing?

    • Deleting the vocab
    • Stopping pad positions from getting softmax weight so they do not mix into real tokens
    • Raising temperature
    • Training LoRA

23. Why Transformers

    23. What is a main training advantage of transformers over classic RNNs?

    • They do not need a tokenizer
    • They can score all prefix positions in parallel instead of stepping token by token
    • They never use softmax
    • They cannot overfit

24. The Stack an Agent Actually Runs

    24. What is the main thing you control in a hosted transformer agent?

    • The number of heads in block 17
    • The token sequence, the sampler, and the stop/schema rules
    • The embedding table rows by hand each turn
    • LayerNorm eps