Job listings for AI training work use terms like “SFT”, “RLHF”, “rubric points” and “golden solution” without explaining them. Here is what each one means, and what you would be doing.
One idea sits under all of it: two people given the same task should grade a model the same way. Nearly every rule below exists to make that true.
The two ways models get trained
Almost every task you will be given belongs to one of two methods. Your project brief will say which.
Supervised fine-tuning
- Project: Writes the prompt
- You: Write the one correct answer, in the exact format asked for
- Model: Imitates it, and copies the format too
No rubric. The answer is the target.
Reinforcement learning from human feedback
- Project: Writes the prompt
- Model: Returns two or more answers
- You: Score each one against the rubric, point by point
- Model: Learns to prefer whatever scored higher
The rubric is the target. It defines what “better” means.
Supervised fine-tuning (SFT)
SFT teaches a model to imitate correct answers. The project supplies prompts, you supply the right answer to each, and the model learns from the pairs.
It learns the format as much as the content. Say the task is pulling dates out of contracts:
- Prompt: “Return the start date as YYYY-MM-DD. The agreement begins on 3 March 2027.”
- What the model gave: “The agreement begins on 3 March 2027, so the start date is March 3rd, 2027.”
- The answer you write:
2027-03-03
The model was never wrong about the date. It was wrong about the shape of the answer. After enough pairs like that one, it stops restating the question and returns the bare date, in the format asked for, every time.
SFT needs no rubric, because the answer you write is the target. That is why SFT work is judged on how closely your answer matches the shape the brief asked for.
Reinforcement learning from human feedback (RLHF)
RLHF teaches a model to prefer the kind of answer a person rates higher. There is no single correct answer to copy, so you score what the model produced.
Here the rubric is not optional. It defines what “better” means, and it is what the model optimises against.
Rubric points: how a judgement becomes a score
A rubric point is one pass/fail check against the model’s answer. A rubric is a list of them.
A good rubric point:
- Covers one thing. Two ideas in one point mean a failure tells you nothing about which half broke.
- Is objective. Someone else reading it should reach your verdict without guessing your intent.
- Is pass or fail. Not a 1-to-10 scale, not “mostly”.
A rubric in use
Say the prompt is this:
Rewrite this cooking step for a beginner. Use fewer than 40 words. Keep both actions, in the order given. Add nothing the original does not say.
Original: Sear the beef on both sides, then transfer the pan to the oven.
Four rubric points fall out of it, and each is a single check anyone can run:
| Rubric point | Passes if |
|---|---|
| Length | Fewer than 40 words |
| Both actions | Mentions searing and the oven |
| Order | Searing comes before the oven |
| Nothing added | No step or ingredient the original never named |
Now score two rewrites.
Rewrite A: “Put the pan in the oven, then brown the beef. Season generously first.”
Length passes and both actions are there. But it flips the order, which would ruin the dish, and invents seasoning the original never mentioned. 2 of 4.
Rewrite B: “Brown the beef on both sides in the pan. Then move the whole pan into the oven.”
Short, both actions, right order, nothing added. 4 of 4.
No part of that scoring needed your taste in food. That is the point: another annotator given the same rubric lands on 2 and 4 too. The model then runs the prompt again and again, keeps what scores 4 of 4, and learns to respect the constraint without ever seeing the rubric.
Golden solutions
A golden solution is the reference answer you write, with your reasoning shown, for whoever reviews your work.
Its job is to let another person follow your logic without asking you what you meant. Write it the way you would want to receive one: structured and readable by someone with no other context.
What makes a good prompt
If your task includes writing the prompts, the best prompt is the one that fits the project’s scope, not the most inventive one. That means:
- Staying inside the scope the brief defines
- Stating the outcome you want, specifically
- Writing your assumptions down instead of leaving them implied
- Defining the structure the answer should take
The cooking prompt above does all four. Every rubric point is checkable because the prompt stated the rule it checks.
Reviewing the model’s reasoning
Some tasks ask you to read the model’s chain of thought, not just its final answer. Two habits matter:
- Read with an open mind. Current models sometimes reason better than the answer set you were handed. If the model’s logic holds and stays in scope, update your set rather than marking it wrong on principle.
- Find the first break. Mark where the reasoning first goes wrong, note the fix in one line, and move on. You are locating the break, not rewriting the whole chain.
Where these terms show up
Screening tasks, guideline documents and the interview itself. Knowing what a platform means by “rubric point” or “golden solution” before you are asked to write one saves you guessing during a timed assessment.
Next steps
- Read how to get started with AI training work for the platforms and the application process
- Check the platform pages to see which ones run this kind of task work
- Browse open roles across every platform this site tracks