RLHF and preference data
Prompts, candidate responses and human judgments: rankings, ratings, critiques and rewritten ideal answers.
- Measured in
- Comparisons or rated prompts, by domain and rater expertise.
- Who buys
- Foundation model labs and companies fine-tuning models for specialist work.
- Where it comes from
- Human data vendors, professional networks, firms with expert staff, and research groups.
- Rights to clear
- Rater contracts assigning rights in their judgments, and terms of the model that generated the responses.
- Privacy
- Low if prompts are synthetic; rater identities kept private.
- Pricing signal
- Usually priced per task or per expert hour. Expert domains cost far more than general rating.
- Preference data records which of two or more model responses a person judged better, and why.
- Labs use it to align models through reinforcement learning from human feedback and related methods.
- Expert raters in law, medicine, finance and engineering are in short supply, which makes domain preference data the most valuable kind.
- Rater qualifications, written guidelines and agreement rates are what buyers check first.
How preference data is made
A rater sees a prompt and two or more responses, picks the better one, and often explains why or writes a better answer. Thousands of these judgments teach a model what people consider helpful, accurate and safe. Critiques and rewrites from real experts teach the most.
What buyers check
- Rater qualifications. Verified credentials for expert domains.
- Guidelines. A written rubric so judgments are consistent.
- Agreement. How often independent raters agree; low agreement signals a vague task.
- Freshness. Responses from current models, since judgments on outdated outputs teach less.
For professional firms
A law, accounting or engineering firm with spare expert capacity can supply preference data or demonstrations, usually as a services contract. Confidential client material must never be used in prompts. See expert-annotated data.
Questions
What is RLHF?
Reinforcement learning from human feedback: training a model to produce responses people prefer, using human judgments as the signal. See the glossary.
Can my firm's experts supply preference data?
Yes, as a services engagement. The buyer usually supplies prompts and responses, and your experts judge or rewrite them under agreed guidelines.
Related guides
Related data types
Need rlhf and preference data, or hold it?
Tell us what you need or what you hold. We check rights and privacy first, then introduce vetted counterparties. You license directly with each other.
- We reply to every enquiry within two working days.
- No upfront fees for sellers
- We never hold your raw data