Meta Description: Compare the top RLHF and LLM training data companies in 2027, from expert networks to evaluation platforms, for SFT, preference data, red teaming and evals.
Large language models are only as good as the human judgment that shapes them after pretraining. Reinforcement learning from human feedback (RLHF), supervised fine-tuning (SFT), expert reasoning traces, red teaming and model evaluation all depend on people who can write a better answer, rank two answers correctly, or spot where a model’s reasoning breaks down. Going into 2027, that work has moved well beyond simple preference clicks. Model builders now ask for PhD-level math solutions, production-grade code, multi-step agent trajectories, verifiable rubrics and reinforcement learning environments that mirror real professional work.
This list is written for ML engineers, AI research leads, AI product managers and founders who need to buy that kind of data, whether they are training a frontier model or fine-tuning an open-weight model for a specific domain. If your project is closer to classic labeling for computer vision, documents or audio, our list of data annotation and labeling companies is a better starting point.
The market has changed quickly. Meta’s 2025 investment in Scale AI pushed several labs to spread their spending across more vendors, and a new group of expert networks grew into businesses with hundreds of millions of dollars in annual revenue within a year or two. Security has also become a buying criterion after a widely reported 2026 breach at one major vendor. The 10 companies below were chosen because LLM and generative AI human data is central to their business, and every fact listed was checked against the company’s own site or reputable news coverage.
RLHF And LLM Training Data Companies At A Glance
Company
Headquarters
Best For
Core Services
Surge AI
San Francisco, CA
Frontier labs needing premium RLHF data
RLHF, human evaluation, safety and alignment data, benchmarks
Scale AI
San Francisco, CA
Large programs and government or regulated work
GenAI data engine, red teaming, evaluation, SEAL benchmarks
Mercor
San Francisco, CA
Professional-grade expert data in law, medicine and finance
Expert human data, evaluations, RL environments, benchmarks
Turing
San Francisco, CA
Coding and reasoning data at scale
SFT, RLHF, preference pairs, evals, coding data
Invisible Technologies
Offices in New York, San Francisco, Washington, D.C. and London
Labs and enterprises that want managed expert operations
Multilingual and safety data with strong compliance
Post-training data, red teaming, agent evals, RL tasks
Prolific
United Kingdom (founded at Oxford)
Representative human feedback and evals
Human Feedback API, preference data, safety evals, HUMAINE
Best RLHF And LLM Training Data Companies In 2027
1. Surge AI — Bootstrapped RLHF Specialist With Over $1 Billion In Annual Revenue
Surge AI was founded in 2020 by Edwin Chen, a former Google and Meta engineer, and is based in San Francisco. It is one of the few large AI data companies built without venture funding: reporting from SiliconANGLE and others in 2025 described the company as bootstrapped and profitable, with annual revenue above $1 billion. Wikipedia’s company profile, citing press reports, puts 2024 revenue at $1.2 billion with roughly 110 employees.
Surge built its reputation on RLHF. Its RLHF page lists training instruction-following and RLHF models, human evaluation of model quality, alignment and safety work, tool use, programming assistants and STEM labeling, delivered through either its platform or a managed expert team. The company has published a case study on how Anthropic used its platform for RLHF data, and its site lists Anthropic, Google, Microsoft, Cohere and AI2 among the organizations it has worked with. More recently Surge has put weight behind evaluation, publishing a series of benchmarks such as Hemingway-bench, Riemann-bench and the DAYJOB set covering tax, healthcare and finance tasks.
The workforce model is a curated contractor network with a small in-house team managing quality, rather than an open crowd. Surge suits frontier labs and well-funded model teams that need high-quality preference and demonstration data and are willing to pay premium rates for it. Smaller buyers should expect a sales process aimed at large, long-running programs.
2. Scale AI — Founded In 2016 With FedRAMP High Authorization For Regulated GenAI Work
Scale AI was founded in 2016 by Alexandr Wang and Lucy Guo through Y Combinator and is headquartered in San Francisco. In June 2025 Meta agreed to take a 49% non-voting stake in the company, reported at about $14.3 billion, and Wang left to join Meta. Scale remains an independent company, and in July 2026 it named former Google Cloud COO Francis deSouza as CEO, according to Axios. The company has said its enterprise and government applications business is expected to overtake its data business within 18 months.
For LLM work, Scale’s GenAI data engine produces training and evaluation datasets, and the company describes itself as an industry leader in GenAI red teaming. Its Safety, Evaluation and Alignment Lab (SEAL) runs private benchmarks and leaderboards, and in March 2026 it became the basis of a broader research division, Scale Labs. Contributors are sourced through Scale’s Outlier platform for generative AI tasks, and the company says 25% of its contributors hold advanced degrees. Scale’s homepage claims that 90% of leading generative AI model builders are powered by its data.
Scale lists ISO 27001, SOC 2, DoD IL4 provisional authorization and FedRAMP High authorization, which makes it one of the few options for defense and public sector model programs. It is best for very large programs and regulated buyers. Labs that compete directly with Meta should review the ownership structure and data separation terms before signing.
3. Mercor — Network Of More Than 5 Million Experts At A $93 Average Hourly Rate
Mercor was founded in 2023 by Brendan Foody, Adarsh Hiremath and Surya Midha and is based in San Francisco. It started as a marketplace for matching freelance engineers with companies, then moved into supplying professional experts to AI labs. A $350 million Series C in October 2025 valued the company at $10 billion. Its customers have included OpenAI and Anthropic.
Mercor’s model is a large professional network rather than a generalist crowd. The company says it has more than 5 million experts, including physicians, lawyers, engineers and consultants, with an average contracted rate of $93 per hour. Its services cover human data and evaluations, reinforcement learning environments built from real professional work, and public benchmarks built with partners, such as APEX-Accounting, APEX-SWE and the Harvey legal benchmark with 1,749 tasks across 24 practice areas. It also offers a data monetization product that lets enterprises license workflow data with automated anonymization.
Buyers should know about the March 2026 security incident. Mercor disclosed a breach linked to a compromise of the open-source LiteLLM package, and TechCrunch reported that Meta paused its work with the company while others reviewed their exposure. The company has since continued operating and growing. Mercor is a strong fit for teams that need credentialed professionals for domain evaluations and expert SFT data, provided their security review covers vendor tooling and contractor data handling.
4. Turing — 4 Million Vetted Engineers Supplying Coding And Reasoning Data
Turing was founded in 2018 by Jonathan Siddharth and Vijay Krishnan and lists San Francisco as its location on its company page. It began as a platform for hiring remote software engineers and now works with frontier labs as a research accelerator for post-training. In March 2025 it raised a $111 million Series E at a $2.2 billion valuation, and TechCrunch reported at the time that the company was profitable and that OpenAI was among the labs it supplied with code data.
The core asset is what Turing calls a talent cloud of more than 4 million AI-vetted software engineers, data scientists and STEM experts. That makes it a natural choice for coding data, which labs have found useful for improving reasoning as well as programming. Turing says its ALAN platform generates human and synthetic data and supports evals, supervised fine-tuning, reinforcement learning, RLHF, preference pair generation, benchmarking and pre and post-training data capture. Its stated focus areas with labs include reasoning, coding, agentic behavior and multilinguality.
Turing also builds AI applications for enterprises, so the same company can supply training data to a lab and deploy models inside a business. It suits model builders whose priority is code, software engineering trajectories and STEM reasoning, and enterprises that want one partner for both data generation and model evaluation on their own workloads.
5. Invisible Technologies — 24,000 Vetted Experts And Claims Of Training Over 80% Of Leading Models
Invisible Technologies was founded in 2015 by Francis Pedraza and has offices in New York, San Francisco, Washington, D.C. and London. Matthew Fitzpatrick, formerly global head of QuantumBlack Labs at McKinsey, became CEO in January 2025. In September 2025 the company raised $100 million led by Vanara Capital, which Bloomberg reported valued it at more than $2 billion, and SiliconANGLE reported 2024 revenue of $134 million.
Invisible combines a full-time team of about 350 people with an expert marketplace. Its homepage lists more than 24,000 vetted experts matched by skill, language and domain, and the company says it has trained foundation models for more than 80% of the world’s leading AI model providers, naming Cohere, Microsoft and AWS in its funding announcement. For LLM work it supplies RLHF data, expert annotation, stress testing of AI systems and reinforcement learning environments verified by domain experts. Its Synapse product scores AI systems in realistic scenarios on accuracy, safety, tone and domain fit, and Neuron helps manage training datasets.
The company positions itself as a managed operations partner rather than a self-serve tool, and it increasingly sells to enterprises that want to fine-tune and deploy models on their own processes. It is a good fit for buyers who want a vendor to own workflow design, expert sourcing and quality control, and who value one team that can move from training data into production automation.
6. Snorkel AI — Stanford AI Lab Spinout Valued At $3.5 Billion In September 2026
Snorkel AI was founded in 2019 out of the Stanford AI Lab and is headquartered in Redwood City, California, with co-founder Alex Ratner as CEO. The company first became known for programmatic labeling software, then expanded into expert data for frontier models. In September 2026 it announced a $350 million Series E led by Insight Partners and S32 at a $3.5 billion valuation, and said its annual recurring revenue had passed $375 million.
Snorkel’s Expert Data-as-a-Service pairs a network of paid subject matter experts with its programmatic platform for labeling and quality control. Its site lists expert demonstrations and reasoning traces, tool-use demonstrations, preference labels and rankings, trajectory QA, helpful and harmless ranking, rubrics with verifiable outcomes such as unit tests and deterministic graders, and environments that include repo and CLI tools, browser and GUI harnesses and simulated multi-step workflows. It also publishes leaderboards for agentic coding, computer use and software engineering, and has committed $30 million to an Open Benchmarks Grants program.
The research background shows in its approach: Snorkel leans on programmatic checks to measure and filter human data instead of relying on reviewer layers alone. Its site displays SOC 2 and HIPAA compliance. It suits labs and enterprises that want rigorous, measurable quality control on expert data, especially for agentic and coding tasks, and enterprises such as banks and insurers that need specialized models evaluated against their own workflows.
7. micro1 — AI Interviewer Zara Vets Experts For A $500 Million Run-Rate Business
micro1 is a San Francisco company led by founder and CEO Ali Ansari. It has grown faster than almost any vendor on this list: TechCrunch reported about $50 million in annual recurring revenue and a $500 million valuation in September 2025, and in August 2026 it reported a $500 million gross annual run rate. Forbes reported in September 2026 that the company was raising more than $100 million at a $4 billion valuation.
micro1’s differentiator is how it recruits. Its in-house AI interviewer, Zara, screens and vets applicants before they join the expert pool, which lets the company add domain experts such as engineers, doctors and lawyers quickly. TechCrunch reported that the network has included professors from Stanford and Harvard. The company describes its offering as expert human data, real-world training environments and contextual evaluations. Its Realm product provides reinforcement learning environments that mirror real-world scenarios for agentic tasks and frontier evaluations for AI labs, Cortex is a contextual evaluation platform for agents in production, and a robotics line supplies real-world data for embodied systems.
micro1 also publishes research benchmarks, including RealmPathology, which tests whether frontier models can read pathology reports. Reported customers include frontier labs, Microsoft, Amazon and robotics companies such as 1X. It is a good fit for buyers who need to stand up a specialized expert team quickly, particularly for agent and domain evaluation work.
8. Labelbox — Alignerr Network Of 2.6 Million Skill-Tested Experts Across 200+ Domains
Labelbox was founded in 2018 and is headquartered in San Francisco’s Mission District, with co-founder Manu Sharma as CEO. It began as a labeling software platform and now combines that software with a managed expert network. In October 2026 it closed a Series E of more than $100 million, including equity and a credit facility, led by Infinitum Partners with participation from Andreessen Horowitz, Wipro Ventures and B Capital.
The expert side runs through Alignerr, which Labelbox says includes more than 2.6 million skill-tested experts across more than 200 domains, including law, finance, medicine and science. These experts write rubrics, preferences and graded assessments for post-training. Horizon, its post-training product, provides reinforcement learning environments and evals covering reasoning, tool use and computer use, along with WorldSim, which builds simulated enterprise apps and business data for agents. Labelbox says more than 90% of leading US AI labs post-train and evaluate on its platform, and it cites work on Meta’s GIM benchmark of 820 expert-authored problems.
Labelbox lists SOC 2 Type II, GDPR, CCPA, NIST 800-171 and PCI DSS compliance. The platform-plus-network model suits teams that want to run some projects in-house with their own reviewers while sourcing experts for others, all on the same tooling. It is one of the better options for buyers who want visibility into the labeling workflow rather than receiving finished datasets only.
9. Toloka — Amsterdam Data Partner Covering 100+ Countries And 40+ Languages
Toloka was founded in 2014 by Olga Megorskaya, who remains CEO, and is headquartered in Amsterdam with offices in the US, Israel, Switzerland and Serbia. It moved from a crowdsourcing and microtask platform to a provider of training data for large language models. In May 2025 it raised $72 million in a round led by Bezos Expeditions, with Shopify CTO Mikhail Parakhin becoming executive chairman. Nebius Group keeps a majority economic stake but gave up majority voting control, giving Toloka more independence.
Toloka sources domain experts through its Mindrift platform and says it works with more than 6,000 active contributors across 90+ domains, with over 70% holding advanced degrees, plus a wider pool from more than 100 countries speaking over 40 languages. Its LLM services include domain-specific demonstrations and preference data, step-by-step reasoning chains, reinforcement learning tasks with built-in verification, RL gyms, agent trajectory demonstrations, expert evaluation of agents across tool-use workflows, and red teaming for prompt injection and policy compliance. It also supplies programming data for coding assistants.
Toloka names Anthropic, Amazon and Microsoft as clients and holds ISO 27001, ISO 27701 and SOC 2 Type II, with private and on-premises storage options. That combination of multilingual reach, safety testing and European data governance makes it a strong choice for teams building models for non-English markets and for buyers with strict privacy requirements.
10. Prolific — 50+ Identity And Behavior Checks With A 55% Pass Rate
Prolific was co-founded at Oxford University in 2014 and is led by co-founder and CEO Phelim Bradley. It started as a platform for academic researchers to recruit study participants and has become a major source of human feedback for AI training and evaluation. In a case study on its site, Ai2 describes cutting human data collection from weeks to hours.
Prolific’s strength is the quality and representativeness of its participant pool rather than raw headcount. The company says joining requires passing more than 50 identity and behavioral checks, with a 55% pass rate, followed by periodic video selfie checks and real-time behavioral monitoring. In 2025 its pool completed more than 380,000 studies and tasks and contributed over 8 million hours, and it now has more than 11,000 AI annotation specialists plus verified expert pools in healthcare, programming, STEM, accounting and languages. Services include a Human Feedback API, an AI Task Builder, agentic evals and benchmarks, methodology design and full project execution.
Prolific also runs HUMAINE, a human-centered model leaderboard backed by about 27,000 evaluators that can be stratified to match US and UK census demographics. It is the best fit on this list for preference tuning, safety evaluations and benchmarks where who gives the feedback matters as much as how much feedback you collect, and for teams that want self-serve access through an API.
How To Choose An RLHF And LLM Training Data Company
How Does The Vendor Vet Its Experts?
Ask for the exact screening path a contributor follows before touching your project: credential checks, skill tests in your domain, identity verification and trial tasks with known answers. A vendor that claims PhD-level math data should be able to show the test its contributors passed and the share of applicants who failed it. Also ask how experts are matched to tasks after onboarding, because a large network matters less than the number of qualified people available for your specific domain and language in a given week.
How Is Quality Measured, Including Inter-Annotator Agreement?
Preference data is only useful if different qualified reviewers would make similar judgments. Request inter-annotator agreement figures from past projects, how disagreements are resolved, and what share of tasks receive a second review. For reasoning and coding data, look for vendors that use verifiable checks such as unit tests, deterministic graders or gold-standard answers rather than relying only on human spot checks. Agree on quality metrics and acceptance thresholds in the contract, and run a paid pilot that you score yourself before committing to volume.
Who Owns The Data And How Is It Secured?
Your prompts, model outputs and rubrics reveal a great deal about your roadmap. Confirm in writing that you own all delivered data and that it will not be reused for other clients. Check certifications such as SOC 2 Type II and ISO 27001, but also ask how contractor devices are managed, which third-party tools touch your data, and how quickly the vendor will notify you of an incident. The 2026 breach traced to a compromised open-source package showed that a vendor’s software supply chain is part of your risk surface, and buyers competing with a vendor’s investors should review ownership and data separation terms.
Can The Vendor Meet Your Throughput And Turnaround Needs?
Post-training runs often need thousands of high-quality comparisons or demonstrations within days. Ask how long it takes to staff a new project, how many qualified contributors are active in your domain today, and what weekly volume the vendor has sustained for similar work. Check time zone and language coverage if you need around-the-clock delivery or non-English data, and ask how the vendor handles surges without lowering its quality bar.
What Pricing Model Fits Your Project?
Vendors price by task, by hour of expert time, or through managed programs with a fixed monthly fee. Hourly pricing is common for specialist work, where rates for credentialed experts can run well above $100 an hour, while per-task pricing suits high-volume, well-defined comparisons. Ask what is included in the price: project design, reviewer layers, rework on rejected items and tooling access often vary between vendors. Comparing cost per accepted, high-quality item gives a fairer picture than comparing headline rates.
Conclusion
The RLHF and LLM training data market in 2027 is defined by expertise and verification rather than headcount. Surge AI, Scale AI and Mercor remain the vendors most frontier labs already know, while Turing, Invisible Technologies, Snorkel AI and micro1 have each built large businesses around coding data, managed expert operations, programmatic quality control or fast expert recruiting. Labelbox, Toloka and Prolific give buyers more control through platforms, multilingual coverage and representative human feedback. Define the skills, quality metrics and security terms your data requires, then run a scored pilot with two or three of these vendors before committing to a long-term program.
