Blog

Working notes on buying human feedback: what each training method asks raters to do, how preference pairs are collected, what rater pay signals and how to write a rubric that two people apply the same way.

  • Rubrics
    4 min read

    How to write rubrics for a human feedback project

    A working rubric names a few criteria in priority order, defines every point on its scale with an example, and settles the common close calls in writing. Test it on a pilot, look at where raters disagree, and change the rubric (with a new version number) before you scale.

  • Data
    4 min read

    Preference data: how pairwise comparisons are collected

    Preference data is collected by showing a rater one prompt and two model responses in random order, and asking which is better against a written rubric. Most projects also record how much better, a short reason and the rater's ID, so the pairs can train a reward model or be used directly with DPO.

  • Pricing
    4 min read

    Rater pay and quality: what the rates tell you

    Rater pay tells you who is likely doing your task and how they are pushed to work. The few published figures run from Prolific's $8 hourly minimum for participants to the up-to-$150 hourly figure Alignerr states for experts. Ask any vendor what raters earn, whether pay is per task or per hour, and how quality affects it.

  • Methods
    4 min read

    RLHF vs DPO vs RLAIF: what are you paying people to do?

    RLHF and DPO both need the same raw material from people: judgments of which response is better. RLHF turns those judgments into a reward model and then runs reinforcement learning; DPO trains on the pairs directly. RLAIF swaps most of the human labels for a model's, so the people you pay write principles, build test sets and audit the machine's labels instead.

Get a quote

Sketch the feedback work in about a minute. What comes back is one vendor name, not a round of cold calls.

RLHF Companies introduces teams to vendors, and the vendor can pay us for that introduction. Sending a brief costs your team nothing.

Keep prompts, model outputs, datasets and samples out of this form; describe the work in words instead.

Part 1 of 3

Your company's location

For a team outside the US, use the ZIP code of your US office or entity.

Get a quote