All learning paths Learn See Try Make Write the answer others are rated against What this lesson is about Write the reference answer itself, not a judgment of somebody else's. It settles the case asked, shows its steps, and refuses where refusing is right.
You are writing the gold reply for a shopkeeper. He was asked whether cut cloth can go back. What do you write first?
A Write the fullest reply you can, covering every kind of return the shop allows. B Write what a good shop ought to do for this buyer, and leave the shop's rule out. C The gold reply settles the case that was asked, not the one you wish had come.
Check my recall Why this matters Many projects buy the ideal answer itself, not a rating of somebody else's. In the published work on training with human feedback, people wrote the wanted answer first. The rating step came after it. A soft reference answer makes every rating built on it soft.
Write the reference answer that every later rating is measured against.
Many projects buy the ideal answer itself, not a rating of somebody else's. In the published work on training with human feedback, people wrote the wanted answer first. The rating step came after it. A soft reference answer makes every rating built on it soft.
The gold reply settles the case that was asked, not the one you wish had come. Show every step a rater must check, including the working you do in your head. Where the honest answer is no, the gold reply says no and says on what. Task A cloth wholesaler's customer writes about a return. He bought forty metres on 2 September, and six were cut to his measurement. He wants to send back eight.
Weak approach Replies that unused goods come back within fifteen days, and asks for the bill. Adds a line about the shop's care for its buyers. Cut cloth is never mentioned.
Stronger approach Says the fifteen days run from 2 September, so 17 September is the last day. Says the six cut metres cannot come back, and why. Then gives what is still open: the uncut metres, exchanged against the same bill. Ends with what the customer must bring.
Why the stronger approach works Both replies are polite and neither says anything false. Only one decides this customer's case, and a rater can point at the line that did it.
Try a changed situation The gold answer you write now is a Class 8 maths question about a discount. What must it carry?
Make something yourself Take a question customers really ask in a trade you know. Write the gold reply for it. Then write three lines on what a rater should take marks off for.
Lesson 8 of 9 on AI trainer essentials next Run a check round and reset the scale Part 2 ends in a work sample you can send
Read the primary or official source: Ouyang et al., Training language models to follow instructions (arXiv)