DataPacts

Expert Demonstrations and SFT Data: How Labs Buy Written Expertise

Checked

At a glance

Expert demonstrations, also called SFT data, are answers, solutions and step sequences written by people to show an AI model what a good response looks like, and AI labs obtain them from human data companies, under services agreements or by licensing archives of existing writing.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Fifteen companies state on their own pages that they supply this kind of work, and each engages its writers through a named entity on its own terms.

The record is detailed on who writes and on whose terms, and thin on price.

Contributors are paid per item or per hour (Shipd lists USD 80 to 220 per item, checked 2026-11-09), the one public services contract gives the supplier ownership of what it delivers, and no company publishes what a lab pays per demonstration.

This page sets out the written expertise inside AI training data licensing, as of 2026-11-16.

What is SFT data? Demonstrations written by people

SFT data is a set of prompts paired with the answers a person wrote for them, used for supervised fine-tuning, the training step in which a model learns to imitate those answers.

The two published volume records are small by web standards. OpenAI's InstructGPT paper (arXiv 2203.02155, March 2022) used about 13,000 SFT training prompts written with about 40 contractors, and Meta's Llama 2 paper (arXiv 2307.09288, 2023-07-19) reports 27,540 SFT annotations.

"Expert" means the writer holds the domain knowledge that the answer needs, such as a software engineer for a coding task or a physician for a clinical question, as platforms list the roles.

Handshake AI lists software engineering and domain tasks for US-based fellows, and Mindrift lists STEM, coding, medicine and consulting. The writer's credentials, not the volume, are what the buyer pays for.

SFT is the demonstration input of the pipeline described under what reinforcement learning from human feedback needs, while ranked comparisons of model answers are a different product with its own suppliers.

Three forms of demonstration data: written answers, worked solutions and trajectories

Companies describe the demonstrations they collect in three forms: written responses to prompts, worked solutions to problems and code tasks, and recorded sequences of steps an agent or person takes on a computer. The table sorts the companies by the wording on their own pages, as of 2026-11-16 (Shipd categories) and 2026-11-09 (all other pages).

Form What a person produces Companies whose pages name it (their wording) Source
Written expert responses a prompt and the answer a model should give AfterQuery ("expert SFT ... data"); Datacurve ("SFT data"); Outlier ("prompt writing", "domain expert tasks"); Prolific ("SFT data") afterquery.com; datacurve.ai/products; outlier.ai; prolific.com/pricing
Worked solutions and code step-by-step solutions to problems, coding challenges, software tasks Alignerr ("coding challenges"); Handshake AI Fellowship ("software engineering", "domain tasks"); Mindrift ("STEM", "coding"); Shipd ("STEM problem solutions"; quests in Software Engineering, Machine Learning, Data Science) alignerr.com; joinhandshake.com/ai; mindrift.ai; shipd.ai
Agent and computer-use trajectories a recorded sequence of steps taken to complete a task on a computer AfterQuery ("computer-use data"); Datacurve ("agent trajectories"); micro1 (licenses approved trajectories from companies' own agents) afterquery.com; datacurve.ai/products; micro1.ai/data-partnerships

Trajectories are also licensed from companies that run their own agents, as micro1 states it. Their buyers and terms are covered under computer-use and agent trajectory data and are not repeated here. Physical demonstrations recorded for robots are a separate market and are not covered on this page.

Who supplies expert demonstrations: 15 companies and their contract entities

Fifteen companies state on their own pages that they supply SFT data, written expert tasks or post-training data, or run an expert network for AI training, and each engages its writers through a named entity on its own terms. The table is alphabetical, as of 2026-11-16, with checks of 2026-11-09 unless marked.

  • "States" means the company's own wording on the check date.
  • "Claim" means a figure as published, never a measured earning.
  • "Not published" means that no page read names it.
  • Inclusion means own-page wording; the order is alphabetical and carries no ranking.
Company Contracting or legal entity on record What its own pages state it supplies or asks writers for Dated claim (scale or pay) Source
AfterQuery legal name not published (terms name only "AfterQuery", Delaware law) expert SFT, RL and computer-use data and agent environments expert network of "nearly 100,000" (2026-04-09) afterquery.com
Alignerr ALIGNERR LLC ("powered by Labelbox") expert reasoning review, coding challenges, rubric grading listings USD 40 to 120/hr; average "$80/hr" (2026-11-09) alignerr.com
Datacurve (Shipd) Datacurve AI, Inc. SFT data, agent trajectories, datasets, benchmarks; Shipd quests incl. STEM problem solutions Shipd bounties USD 80 to 220 per item (2026-11-09) datacurve.ai; shipd.ai
Deccan AI Deccan.ai, Inc. post-training data, evaluation and RL environments through an expert network not published deccan.ai
Handshake AI Fellowship Stryder Corp. (d/b/a Handshake); Handshake AI Solutions, LLC expert evaluation, software engineering, domain tasks; US-based fellows Software Engineer up to USD 150/hr; page meta "Earn up to $100/hr" (2026-11-09) joinhandshake.com/ai
Innodata Innodata Inc. training and post-training data and evaluation for AI labs and Big Tech two largest customers about 43 % and 31 % of Q3 2026 revenue (10-Q, 2026-09-30) innodata.com; sec.gov
Invisible Technologies Invisible Technologies, Inc. AI training data and RL environments; expert network Meridial "24,000+" Meridial experts (2026-11-09) invisibletech.ai
Mercor Mercor.io Corporation expert-contributor marketplace for AI labs "nearly five million" experts (2026-06-25); home page "5M+" (2026-11-09) mercor.com
micro1 micro1 Inc. expert contributors, RL environments, evaluations; licenses approved agent trajectories from companies trajectories "$1-2M annually" to the company (2026-11-09) micro1.ai
Outlier (Scale AI) Smart Ecosystem, Inc. (dba Outlier AI) with Scale AI, Inc. prompt writing, domain expert tasks, response evaluation and ranking pays "either hourly or per task", no rate published (2026-11-09) outlier.ai
Prolific entity not stated on its pages human feedback, preference data, evaluations, SFT data from participants researchers set pay: minimum GBP 6 / USD 8 per hour (help page 2026-07-17) prolific.com
Proximal Proximal AI, Inc. post-training data and coding RL environments under a public services agreement no contributor rate in the record proximal.ai
Snorkel AI Snorkel AI, Inc. datasets and simulated environments delivered as a service contributor terms and pay not in the record snorkel.ai
Surge AI (DataAnnotation) Surge Labs, Inc. coding, STEM, medical, legal, finance and 43 language roles via DataAnnotation general projects "start at $25-$50 per hour" (2026-11-09); one surgehq.ai listing USD 200 to 400/hr (2026-11-16) dataannotation.tech; surgehq.ai
Toloka (Mindrift) Toloka AI BV Mindrift: STEM, coding, medicine, consulting, generalist AI training domain experts "$60-100+/hr"; STEM up to USD 6,400/month (2026-11-09) mindrift.ai; toloka.ai

Three of the 15 name SFT data outright (AfterQuery, Datacurve and Prolific), three describe post-training data (Deccan AI, Innodata and Proximal), and the other nine name expert work or an expert network without naming the output.

Scale claims conflict within companies: Mercor states "nearly five million" experts on one page and "5M+" on another, and as checked on 2026-11-09, Handshake shows "Earn up to $100/hr" beside listings up to USD 150 per hour. Both figures stay side by side. Innodata's customer shares are shares of total revenue, not of SFT revenue.

How labs obtain written expertise: commissioned work, custom builds and licensed archives

Labs obtain demonstrations by three routes on record: commissioning new writing through contributor platforms, ordering custom post-training data under a services agreement, and licensing archives of writing that already exists. The table compares the routes as of 2026-11-16, with terms read on 2026-11-09 unless marked.

Route Who delivers What the lab receives Terms on record Source
Commissioned writing contributor platforms and expert networks (table above) new demonstrations written to the lab's brief contributor assigns work product to the platform (Outlier s.4(b); Shipd s.3.1); lab-side terms not published outlier.ai/legal/terms-of-use; shipd.ai/terms
Custom build under a services agreement Proximal (public MSA) post-training data and environments as Deliverables Proximal owns Deliverables (s.8.1); non-exclusive internal-use license to the customer (s.8.2); reuse for other customers absent customer confidential information proximal.ai/msa (last updated 2026-07-06)
Licensed archive of existing writing Stack Exchange; data holders through Labelbox Prism Q&A pairs; financial and scientific reasoning traces Stack: "continuous access for training and fine-tuning", attribution expected; Prism: exclusive by default, upfront fee plus revenue share stackoverflow.co/data-licensing; prism.labelbox.com/faq

In commissioned work the contributor assigns the work product to the platform and the platform delivers it to a lab client, whose own terms are not published.

The filing record shows how concentrated that client base is at one supplier: Innodata's 10-Q for the period ending 2026-09-30 states that its two largest customers were about 43 % and 31 % of Q3 2026 revenue.

In custom builds the contract is public at one supplier. Proximal's Master Development Services Agreement, last updated 2026-07-06, makes Proximal the owner of the Deliverables (s.8.1), gives the customer a non-exclusive, non-transferable, non-sublicensable, worldwide, royalty-free license for internal business use during the term (s.8.2), and lets Proximal retain learnings and model training (s.8.3).

Invoicing is monthly, undisputed amounts are due within 30 days, and California law governs (s.12.5).

Licensed archives skip the commissioning step. Stack Exchange offers Q&A pairs from 150+ sites under "continuous access for training and fine-tuning" with attribution expected, supplies a sample of 1,000 Q&A pairs and publishes no price, and its agreement with OpenAI announced on 2024-05-06 carries no disclosed value.

Labelbox's Prism program asks data holders for financial and scientific reasoning traces, as covered under Labelbox Prism.

Anthropic's S-1 of 2026-10-26 lists datasets licensed or commissioned from third parties among its training data sources, which names the first and third routes in a filing, without naming a supplier.

Who owns a written demonstration: assignment terms on contributor platforms

A contributor who writes a demonstration usually assigns all rights in it to the platform when it is created or submitted, so the expert keeps no claim to reuse or license the text.

Outlier's terms of use (effective 2026-05-11) have the contributor irrevocably assign all right, title and interest worldwide in the work product (s.4(b)), keep it confidential after termination (s.5), and arbitrate individually with a 30-day opt-out for US residents (s.10).

Shipd's Contributor Terms assign Work Product to Shipd on creation or submission (s.3.1). The entity named differs between documents: the Contributor Terms say Datacurve AI, Inc., while the privacy policy of 2026-06-01 says the Work Product is assigned to "Datacurve, Inc.", and both stay visible here.

Alignerr's terms give it an unlimited, irrevocable, perpetual, royalty-free, sublicensable right to interview material, with no obligation to pay for the interview. What an expert may not bring to a task at all, such as patient records, client secrets or employer code, is set out under expert contributor work by field.

The buyer side mirrors the contributor side: the custom-build route above shows a supplier that keeps ownership of what it builds.

What demonstration data costs: the price points on record

No company publishes what a lab pays per demonstration, so the price points on record are what platforms pay writers, what a participant panel charges on top, and what licensing programs state for existing reasoning data. The table lists six, as of 2026-11-16, each as its stating party's claim.

Price point Who pays whom Value as stated Stating party and date Source
Shipd bounty Datacurve (Shipd) pays the contributor USD 80 to 220 per item Datacurve, 2026-11-09 shipd.ai
Prolific study researcher or AI developer pays participants plus Prolific minimum GBP 6 / USD 8 per hour; fee usually 42.8 % corporate, 33.3 % academic or non-profit Prolific, 2026-07-17 (pay), 2026-11-09 (fee) researcher-help.prolific.com; prolific.com/pricing
Labelbox Prism dataset Labelbox pays the data holder "start around $50K"; enterprise partnerships "can reach $1M+" Labelbox, 2026-11-09 prism.labelbox.com/faq
micro1 agent trajectories micro1 pays the company "$1-2M annually" micro1, 2026-11-09 micro1.ai/data-partnerships
Stack Exchange data licensing AI company pays Stack Exchange not published Stack Exchange, 2026-11-09 stackoverflow.co/data-licensing
OpenAI and Stack Overflow agreement OpenAI and Stack Exchange, Inc. not disclosed Stack Overflow announcement, 2024-05-06 stackoverflow.co/company/press

The rows differ in kind. The Shipd bounty and the Prolific minimum pay writers, the Prism and micro1 figures are program ranges for licensed data and not completed payments, and the last two rows are agreements without a stated value.

Prism adds an upfront fee plus a revenue share, which Labelbox states in the same FAQ.

Hourly claims for writing roles vary by field as of 2026-11-09: Handshake lists up to USD 150 per hour for software engineering, Mindrift states "$60-100+/hr" for domain experts, and DataAnnotation states USD 25 to 50 per hour for general projects. Full rates by platform and field sit under pay for AI training work.

Writing or licensing demonstrations: benefits, risks, downsides and alternatives

Demonstrations pay experts per task or hour and pay data holders per license, and in both cases the text leaves the writer's control once it is delivered. The lists below set out the position of an expert who writes demonstrations or a holder who licenses archives, and every figure named is a claim.

Benefits

  • Paid project work at stated rates for domain writers, as Handshake and Mindrift list them (checked 2026-11-09).
  • Per-item bounties displayed on the public quest page, as Shipd shows them.
  • Licensing existing Q&A or reasoning archives can earn from writing already done, as Stack Exchange and Labelbox Prism describe their programs.

Risks

  • Irrevocable assignment and confidentiality that survives termination, as in Outlier's terms (s.4(b), s.5).
  • Prism's own FAQ states that "Models already trained on the data cannot be unlearned".
  • Identity checks and monitoring on platforms, as Alignerr states them: Persona ID verification and Hubstaff time tracking with periodic screenshots.

Downsides

  • No lab price per demonstration is published, so the value of a piece of work is hard to judge.
  • Work ends with the project, and no platform publishes volumes.
  • Contributor pay has been contested in court: McKinney v. Scale AI (Superior Court of California, County of San Francisco, CGC-24-620481) is settled as of 2026-11-16, with final approval filed on 2026-10-30 for a gross USD 12,500,000, with payments reported on Form 1099.

Alternatives

  • Other human data work, such as rating, covered under AI raters and evaluation data.
  • Licensing company records through a program instead of writing new text.
  • Not writing and not licensing at all.

The decision framework for the holder side, including the case for keeping data, is set out under whether to sell or license data at all.

What the public record does not show about demonstration data

The record leaves four things open: what a lab pays per demonstration, which supplier writes for which lab today, how many demonstrations labs have bought since the Llama 2 paper, and what share of submitted work platforms accept.

  • Price per demonstration or per SFT item: no rate card, order form or filing states one.
  • Supplier-to-lab pairings: platforms name few clients, and the customer statements above are each company's own.
  • Volumes bought after the Llama 2 paper: no lab publishes them.
  • Acceptance rates for submitted work: no platform publishes them.

Where demonstration data sits among human data, trajectories and synthetic data

This page covers written demonstrations; the pages below cover the wider human data market, rating work, contributor lawsuits, physical demonstrations for robots, the case for human data over synthetic data, published prices and how this site is funded.

Frequently asked questions

Is SFT data post-training data?

Suppliers file SFT data with post-training work. AfterQuery lists expert SFT next to RL data, and Proximal, Innodata and Deccan AI describe post-training data. The pages read give no definition beyond that grouping.

How is DPO different from SFT?

SFT learns from written demonstrations, DPO from ranked comparisons. The DPO paper (arXiv 2305.18290) was submitted on 2023-05-29 and trains on ranked comparisons without a separate reward model. The RLHF page linked above compares the methods.

Can I get paid to write SFT data?

Platforms state pay for writing work, as claims. As checked on 2026-11-09, Shipd lists USD 80 to 220 per item, Mindrift "$60-100+/hr" for domain experts and DataAnnotation USD 25 to 50 per hour for general projects. See AI training jobs and platforms compared.

Can a company license its documents as demonstration data?

Two programs state that they take existing writing. Stack Exchange licenses Q&A pairs for training and fine-tuning, and Labelbox Prism asks for reasoning traces. Other terms are listed under company data partnership programs.