CCARF Practice Question

Paper 4 · Question 3

D4 Prompt Engineering & Structured Output · Paper 4

The moderation system classifies hate speech accurately for explicit slurs but misses coded language and dog-whistle terms that human moderators easily recognise. The prompt includes detailed written rules about coded language patterns. What intervention would most improve detection of coded hate speech?

• A.Add 20+ examples covering every known category of coded hate speech to maximise coverage
• B.Add 2–4 few-shot examples of coded hate speech, with reasoning naming the coded language and the targeted group.
• C.Add a comprehensive dictionary of every known coded term and dog-whistle to the system prompt so the classifier can match posts against the full list.
• D.Increase the model's context window so it can consider more of the user's post history for context

Show answer & explanations

Correct answer: B

A: Laundry lists of examples dilute attention and teach pattern-matching rather than generalisation. 2–4 targeted examples with reasoning are more effective than 20 examples without reasoning because they teach the analytical process.

B: Few-shot examples with reasoning tags teach the model the pattern-recognition process, not just individual terms. Showing the reasoning chain — surface meaning, coded meaning, targeted group, contextual signals — enables the model to generalise to new coded terms it has not seen before.

C: A static dictionary becomes outdated immediately as coded language constantly evolves. It also bloats the prompt without teaching the model to recognise the underlying patterns of coded speech.

D: Context window size is not the bottleneck. The model fails to recognise coded language in individual posts, which is a prompt engineering issue, not a context limitation.

Related Prompt Engineering & Structured Output questions

More: Paper 4 · All practice questions