Guide · Human feedback services

Human Feedback Training Services (2026)

Every fine-tune, preference tune and RL run consumes the same raw material: human beings judging model outputs and demonstrating better ones. This page is about that material as a purchased service — what the deliverables actually are, how engagements run, and how the quality is (and is not) controlled.

The five deliverables

Preference pairs. Two (or more) model outputs, one human verdict, ideally with a reason. The input to DPO and RLHF reward modeling — bought by the comparison, priced by the expertise the comparison requires.
SFT demonstrations. An expert writes the response the model should have given. The highest-leverage data in the catalog: a fine-tune is, mechanically, an imitation of its demonstrations. See how they are consumed on the fine-tuning services page.
Graded evaluations. Outputs scored against a rubric — accuracy, instruction-following, tone, safety — producing the measurements that tell a team whether training worked and where it fails.
Safety and red-teaming data. Adversarial prompts and graded responses across harm categories; the dataset that decides what the model refuses and how.
RL environments. Sandboxed tasks — often replicas of real software — in which a model can act and be scored by a verifier. The newest deliverable, and the one growing fastest as agent training replaces chat training.

Human feedback providers compared — the ranked analysis is on the RLHF providers page
Vendor	Model	Founded	Pricing	Best for
Surge AI	Managed service	2020	Not published	Labs that rank data quality above all else for RLHF and reasoning data, and prefer a founder-controlled vendor with no outside investors.
Scale AI	Managed service	2016	Not published	Frontier labs, large enterprises and government programs that need annotation and evaluation run as a managed, high-volume program.
Mercor	Expert network	2023	Not published	AI labs that need credentialed expert judgment — clinical, legal, financial, engineering — at serious scale and speed.
Micro1	Expert network	2022	Not published	Companies that want vetted expert capacity without running their own recruiting funnel, and experts who want clearly posted pay ranges before they apply.
Snorkel AI	Data lab	2019	Not published	Enterprises and labs that want benchmarks, evaluations and expert datasets built with academic-grade rigor, delivered as a service.

Demonstrations & preference data

Surge AI

The specialist: post-training data is the whole catalog, at a reported billion-dollar revenue scale.

Evaluation design

Snorkel AI

When the rubric itself is the hard part — benchmarks and graded evals built as a research service.

Credentialed expert panels

Mercor

Doctors, lawyers, bankers and engineers, supplied hourly, for judgments a generalist pool cannot make.

How an engagement actually runs

Rubric first. The grading criteria are drafted, argued over and frozen before production. Every later quality claim traces back to this document; a provider eager to skip it is telling you how the project ends.
Pilot batch. A small paid batch — hundreds of items — graded by the provider and reviewed by the buyer. This is the procurement: it prices the real cost per accepted item and exposes rubric ambiguity while it is still cheap.
Calibration. Raters are measured against gold items and against each other. Inter-rater agreement is reported, and disagreements in expert domains are adjudicated, not averaged away — the disagreement is often the signal.
Scaled production with QA sampling. Ongoing gold-set checks, drift monitoring and per-rater quality scores. The deliverable includes provenance: who produced each item, under which rubric version.

Staffing the expert tier

The constraint that defines this market is not platform capacity; it is qualified humans. Preference data for a finance model needs people who can tell a good financial analysis from a fluent bad one. Expert networks exist to solve exactly that staffing problem — Micro1, for example, interviews and certifies experts with an AI-led process and matches them to feedback and evaluation projects, publishing per-role pay ranges (roughly $22–$70/hour for generalist roles, $200+/hour for deep specialists, per third-party reviews of its listings) that make the labor economics diligence-able. Companies can register to post projects: register your company with Micro1 →

Disclosure: this is a referral link. If your company registers with Micro1 through it, we may earn a referral commission at no cost to you. It does not change where Micro1 ranks — see our methodology.

Where the data goes

Human feedback is an input, and its buyers are training runs: demonstrations feed the supervised fine-tunes compared on our fine-tuning services page, preference pairs feed DPO and RLHF, graded evals gate every release. That is also why this market belongs in a publication about data annotation: the workforce, the QA mechanics and increasingly the vendors are the same — the judgment is simply worth more per hour than the label ever was. Buyers comparing feedback quotes should use the same unit discipline as annotation: cost per accepted item, after QA, with provenance attached.

Frequently asked questions

What are human feedback training services?

Services that produce the human judgments modern model training runs on: preference pairs (which of two outputs is better, and why), supervised fine-tuning demonstrations (an expert writing the ideal response), rubric-graded evaluations, safety and red-teaming data, and RL environments. Buyers are purchasing graded human judgment as a manufactured input — specified by rubric, produced by vetted people, and quality-controlled like any other production process.

How is human feedback data priced?

By the unit of judgment or by expert time. Generalist preference labeling tracks annotator rates — transparent platforms floor participant pay at $8/hour and skilled generalist work runs well above that — while expert feedback tracks expert rates: $85–$200/hour is CleverX’s published guidance for credentialed specialists, with medical and legal experts above it. Providers rarely publish project pricing; the comparable number across quotes is cost per accepted item after QA, the same discipline as cost per accepted label in annotation.

What is the difference between human feedback services and RLHF providers?

Mostly vocabulary. "RLHF provider" names the technique the data feeds; "human feedback service" names what is actually bought and delivered — demonstrations, preference pairs, graded evals. Our RLHF providers page ranks the sellers; this page explains the merchandise: what the deliverables are, how an engagement runs, and how quality is controlled. The same companies appear in both because they sell the same work.

How does human feedback connect to data annotation?

It is annotation’s top tier. The pipeline is identical — recruit, vet, train on a rubric, produce, QA, deliver — but the unit of work is a judgment about model behavior rather than a label on raw data, and the workforce is experts rather than generalists. Buyers who already purchase annotation will recognize every mechanic; what changes is the price per hour and the cost of a wrong answer.

Who writes the demonstrations in an SFT project?

Whoever the rubric says is qualified: for a coding model, working engineers; for a clinical model, clinicians; for general assistant behavior, trained generalists working to a detailed style and accuracy rubric. Provenance — who wrote each demonstration and what qualified them — is part of the deliverable from serious providers, and buyers should contract for it: a fine-tune inherits its teachers.

Sources

One primary source per provider shown; full lists on each profile.

Facts on this page were drawn from the public sources above, checked November 2026. If you spot an error, tell us at the corrections address in the footer — we fix sourced errors and note the change.

Last updated: November 8, 2026