Practice

14 questions

• DL, NLP and LLMs

  Your neural network's training loss stays flat from the first epoch, and on a second run it becomes NaN after a few hundred steps. Walk me through how you debug it.
  mid10 min

• DL, NLP and LLMs

  Explain self-attention to an engineer who knows RNNs. Why did Transformers largely replace RNNs for language tasks, and what does attention cost?
  mid8 min

• DL, NLP and LLMs

  You must route 50,000 customer emails a month into 12 categories for an airline's service team. You have 8,000 labelled emails. A colleague wants to fine-tune a large language model immediately.
  mid10 min

• DL, NLP and LLMs

  An airline wants an assistant that answers agents' questions about fare rules, baggage policy and rebooking procedures. Policies change monthly.
  senior15 min

• DL, NLP and LLMs

  Explain how LoRA works. For one 4096 x 4096 weight matrix with rank 8, how many trainable parameters does LoRA add, and what are the practical trade-offs compared with full fine-tuning?
  senior12 min

• DL, NLP and LLMs

  Your company is launching LLM-generated summaries of customer-service conversations for supervisors across four languages.
  staff25 min

• DL, NLP and LLMs

  An LLM step that extracts a fare class from customer emails returns different answers when the same email is sent twice. A teammate says "just set temperature to 0 and it is deterministic".
  mid8 min

• DL, NLP and LLMs

  You must fine-tune a transformer classifier, but only 8 sequences of up to 512 tokens fit on one GPU at a time and the recipe you are reproducing used a batch size of 512.
  senior15 min

• DL, NLP and LLMs

  A team fine-tunes an instruction-following model on 20,000 labelled airline customer-service replies. Task accuracy on their evaluation set rises (hypothetical numbers) from 71% to 89%, and they want…
  senior15 min

• DL, NLP and LLMs

  A customer-support routing step calls a large hosted LLM 30 million times a month. An engineer proposes distilling it into a small self-hosted model.
  staff25 min

• DL, NLP and LLMs

  You must classify baggage-damage photos into five damage types with 1,200 labelled images. The photos come from about 400 claims, three photos per claim.
  mid12 min

• DL, NLP and LLMs

  You serve a 7-billion-parameter decoder-only model in 16-bit precision on one 80 GiB GPU for a support assistant with 8,192-token contexts. Throughput is far below the benchmark, and requests queue.
  senior15 min

• DL, NLP and LLMs

  Whiteboard case. An airline's contact centre wants one intent-classification and summarisation capability for customer messages in Turkish, English and a long tail of other languages, replacing…
  staff30 min

• DL, NLP and LLMs

  You compare two prompts for an LLM that extracts the refund reason from customer messages, on the same 200 labelled messages. Prompt A gets 156 right (78%), prompt B gets 164 right (82%).
  senior15 min