During fine-tuning, what do human raters actually do?
AWrite new training data from scratch
BAdd more internet text to the training set
CShow the model examples of good and bad replies
DFix bugs in the model's code directly
Answer & Solution
Correct answer: C. Show the model examples of good and bad replies
Card 3 of this block: in fine-tuning, humans show the model pairs of good replies vs bad replies so the model learns the shape of being helpful. Source: claude_constitution_ocr.md (constitution plays a crucial role in the training process, p.1).
Related questions
A friend asks: what should you do first when an AI answer contains a specific date?Which type of claim carries the highest hallucination risk?What is the main reason AI sometimes produces confident but wrong answers?You told AI your name yesterday. In a new chat today, AI does not know it. Why?AI says an essay has 850 words. The real count is 1003. What does this show?You ask AI for today's cricket score. Why can't it answer reliably?What does AI do well when you ask it to compare two options?Which task is AI best suited for?