AI Training Data Startups to Watch
Patronus AI TeamAs of November 202612 min read
Patronus AI publishes this guide and is one of the companies on it. Entries are grouped by how they work, not ranked. Each fact comes from the company’s own site or announcement, a paper it published, or the profile its entry names. Counts, funding, roles and releases are as of Nov 19, 2026.
Patronus AI builds RL environments and the data inside them. On Jun 25, 2026, it announced a $50M Series B led by Greenfield Partners. AI training data startups sell the examples, rankings, evaluations and environments that AI labs train and test models on. As of November 2026, Patronus AI and Snorkel AI are research labs that publish their methods, Datacurve specializes in coding data, Handshake AI, Mercor, micro1 and Turing run expert networks, and Deccan AI, Invisible, Labelbox, Scale AI and Surge AI sell the full range as data vendors.
Each entry gives what the company sells, who does the work, its funding, its public work and whether it is hiring, where its own pages say so. For RL environments alone, see our guide to RL environment startups.
Snapshot as of
What AI training data is
After pre-training, a model learns skills and behaviors from much less data, mostly demonstrations, preference data and RL environments. Demonstrations are worked examples of a task, from one ideal answer to a long trace of an agent’s steps, and the model learns by imitating them with supervised fine-tuning (SFT). Preference data ranks two or more answers to the same prompt, for a reward model or direct preference optimization. RL environments give an agent tools, a task and a verifier that scores the result.
Most of the companies below run networks of experts, such as physicians, lawyers, bankers and engineers, to write and check this data. Our guide to RL environments covers environments in depth.
How these startups differ
Most of them sell RL environments next to human data. They differ in where the expertise comes from and what else they do.
Research labs that build training data
Patronus AI and Snorkel AI. They publish papers and benchmarks and build data from that research.
Coding data specialists
Datacurve. It focuses on software engineering and data science data.
Expert networks
Handshake AI, Mercor, micro1 and Turing. They recruit large networks of professionals to write and grade data.
Research labs that build training data
These companies call themselves labs and publish the research behind their data, so a buyer can read how they work before a first call.
Patronus AI
Patronus AI is a behavioral research lab for AI. When an agent misses the point or cuts corners, its researchers work out why and then build the training data to correct it.
- What it sells. RL environments with tasks, graders and simulated apps for alignment, coding, knowledge work, science, computer use and more, plus expert pairwise reviews of agent transcripts. It trains open models on its own tasks to measure how much the models learn from them.
- Where its data comes from. Domain experts review every environment before delivery. It also licenses codebases, company records and games from the companies that made them, on non-exclusive terms, through data.patronus.ai.
- Funding. A $50M Series B led by Greenfield Partners, announced Jun 25, 2026, with participation from Lightspeed Venture Partners, Notable Capital, Datadog, Samsung and others, after a $17M Series A led by Notable Capital in May 2024.
- Accepted venues. Main conference papers at ICML 2026, NeurIPS 2026 and ACL 2025, and a paper in Findings of ACL 2026.
- Public work. FinanceBench (2023), TRAIL (2025), GlobeBench (2026) and FigmaTrace, a public dataset of 3,469 human Figma design trajectories, among 20 papers as of Nov 10, 2026.
- People. Patronus AI was founded in 2023 by CEO Anand Kannappan, formerly of Meta Reality Labs, and CTO Rebecca Qian, formerly of FAIR (Meta AI). Its researchers and engineers come from Meta Superintelligence Labs, Amazon AGI, Apple and Google.
- Hiring. Patronus AI lists open roles on its Join page.
Snorkel AI
Snorkel AI calls itself the frontier AI data lab. It builds specialized training data, benchmarks and evaluation environments for high-stakes domains.
- What it sells. Research-grade datasets, evaluation systems and runnable environments, plus the Snorkel Data Series of curriculum-structured datasets with rubrics and difficulty tiers built in.
- Who does the work. An Expert Community of professionals and academics across 1,000+ domains.
- Funding. $350M in a round co-led by Insight Partners and S32, announced Sep 22, 2026.
- Public work. Senior SWE-bench (July 2026), released with researchers at Princeton and UW–Madison, and the MLSys 2026 paper Learning from Less, on RL with verifiable rewards in low data and compute regimes.
- People. Co-founder and CEO Alex Ratner. Snorkel AI was founded out of the Stanford AI Lab in 2019.
- Hiring. Open roles, among them Research Scientist, RL Training.
Coding data specialists
This startup builds data for software engineering and nearby fields such as data science, cybersecurity and machine learning.
Datacurve
Datacurve sells custom data for long-horizon reasoning, software engineering and data science.
- What it sells. RL environments, long-horizon tasks, off-the-shelf datasets, benchmarks, agent trajectories and SFT data, in software engineering, data science, cybersecurity, machine learning and research.
- Who does the work. Engineers on Shipd, its contributor platform, where they solve hard problems for bounties.
- Funding. $17.7M across a $15M Series A led by Chemistry and a $2.7M seed, announced by its CEO on Oct 9, 2025.
- Public work. DeepSWE, a long-horizon coding benchmark introduced in May 2026, with a paper, leaderboard and open code.
- People. Founders Serena Ge, its CEO, and Charley Lee, per its Y Combinator profile (Winter 2024).
- Hiring. Software engineers and research engineers, on its careers page.
Expert networks
These companies run large networks of professionals, vetted at scale, and put them on data projects for AI labs.
Handshake AI
Handshake AI is the AI data business of Handshake, which was founded in 2014. It launched in March 2025 and works with frontier labs to stress-test models against expert workflows.
- What it sells. Expert datasets and evaluations for frontier labs, and RL environments. It builds enterprise RL environments with Fireworks AI.
- Who does the work. Fellows in the Handshake AI Fellowship, 188k+ of them as of Nov 19, 2026, plus 200+ in-house researchers and engineers.
- Public work. The ATLAS benchmark suite, from BankerToolBench (Apr 13, 2026) to ATLAS-Finance (Nov 17, 2026), with open datasets and code.
- Dated deals. It acquired Cleanlab (Jan 22, 2026), Taro (Jan 26, 2026) and Uplimit (Jun 30, 2026).
- People. Co-founder and CEO Garrett Lord, with co-founders Ben Christensen and Scott Ringwelski.
- Hiring. Member of Technical Staff roles in post-training, evals and coding agents.
Mercor
Mercor calls itself an AI data company. It was founded in 2023 and runs a marketplace of experts who make data for AI labs and enterprises.
- What it sells. Benchmarks, evaluation environments, large-scale human datasets and RL environments with tasks and verifiers.
- Who does the work. Physicians, lawyers, engineers, consultants and other experts, 5M+ of them as of Nov 19, 2026.
- Funding. A $750M Series D, announced Nov 10, 2026, with no lead investor named.
- Public work. The APEX benchmarks of professional work, starting Sep 30, 2025, and OWE-Bench (Nov 12, 2026), with open datasets on Hugging Face.
- Dated deal. It acquired Deeptune, an RL environment company, in a deal announced Jul 9, 2026.
- People. Co-founders Brendan Foody and Adarsh Hiremath.
- Hiring. Open roles, among them Research Scientist, APEX Benchmarks.
micro1
micro1 calls itself a data lab that produces data for frontier model training and agent performance.
- What it sells. Realm, RL environments and frontier evaluations built with domain specialists in 100+ fields, Cortex, a platform for evaluating agents in production, and robotics data.
- Who does the work. Domain experts sourced and vetted by Zara, its AI recruiter agent.
- Funding. A Series B led by Dream Ventures, disclosed without an amount in its third quarter update of Oct 12, 2026.
- Public work. Leaderboards such as Dual-Arm Sim and MissionReport, and research papers dated through October 2026.
- People. Founder and CEO Ali Ansari.
- Hiring. Open roles, among them Member of Technical Staff, Realm Lab.
Turing
Turing says it powers frontier AI labs with datasets, RL environments and expert talent, and helps large companies deploy AI agents.
- What it sells. Custom and off-the-shelf datasets, RL environments and human experts in software engineering, enterprise knowledge work and STEM. Its ready-made sets come with tasks, verifiers and traces.
- Who does the work. An expert pool of 5M+ across 250+ domains, as of Nov 19, 2026.
- Funding. A $111M Series E led by Khazanah Nasional Berhad, announced Mar 6, 2025, for $225M raised since its founding in 2018.
- Public work. Training sets such as Turing Terminal-Bench Data (July 2026), and benchmarks such as KernelQuest and Behavior2Code (September 2026).
- People. Co-founder and CEO Jonathan Siddharth and co-founder Vijay Krishnan.
- Hiring. Open roles, among them Senior Research Engineer, Frontier Data.
Full-service data vendors
These companies sell across the range, from human data to evaluations and RL environments, and several also sell AI products or agents to large companies.
Deccan AI
Deccan AI sells training data, RL environments and off-the-shelf datasets to frontier labs, and runs AI agents for enterprises.
- What it sells. Datasets for multimodal, coding, language and physical AI work, private code repositories for training and evaluation, STARK RL environments and Helix evaluations.
- Who does the work. Over 1M contributors in the Deccan experts network, as of Apr 14, 2026, and an in-house team of 100+ quality associates.
- Funding. A $25M Series A led by A91 Partners, with Susquehanna International Group and Prosus Ventures, in a post dated Apr 14, 2026.
- Public work. Benchmarks and leaderboards such as CASTBench, for reference-to-video models, and VoxClash TTS 2.0, for text-to-speech.
- People. Founder Rukesh Reddy.
- Hiring. Open roles, among them ML Researcher, Benchmarks and Evaluation.
Invisible
Invisible Technologies trains AI models for labs and deploys custom AI for large companies. Francis Pedraza founded it in 2015.
- What it sells. RL environments, evaluations, annotation, SFT, RLHF and off-the-shelf training data. Its experts define task logic and design reward rubrics.
- Who does the work. A network of 22M+ experts across 800 domains, as of Nov 19, 2026.
- Funding. $100M in growth funding led by Vanara Capital, announced Sep 16, 2025, for $144M raised in all.
- Public work. Language Atlas (Nov 10, 2026), on multilingual model training, and the paper How Much of Your Data Can Suck? Thresholds for Domain Performance and Emergent Misalignment in LLMs (September 2025).
- People. CEO Matt Fitzpatrick since 2025.
- Hiring. Open roles, among them Research Engineer.
Labelbox
Labelbox builds RL environments that AI labs train in, and Recursion, a platform enterprises use to run agents. It was founded in 2018.
- What it sells. Horizon, RL training gyms and evaluations for reasoning, tool use and computer use, and expert data through Alignerr, such as preference pairs, ranked trajectories and rubric-graded assessments.
- Who does the work. Contributors on Alignerr across 200+ knowledge domains and 40+ countries.
- Funding. More than $100M in Series E financing, equity and a credit facility, led by Infinitum Partners in October 2026.
- Public work. EchoChain (Mar 4, 2026), an audio benchmark for reasoning in full-duplex voice dialogue.
- People. Co-founder and CEO Manu Sharma.
- Hiring. Open roles, among them Forward Deployed Engineer, RL Environments.
Scale AI
Scale AI delivers data, evaluations and outcomes to AI labs, governments and large companies. It was founded in 2016.
- What it sells. Training data, annotation and RLHF, model evaluations and red-teaming, and RL environments built on enterprise systems, MCP servers and desktop computer-use tasks.
- Who does the work. Experts across 370+ fields and 65+ languages, who sign up through Outlier.
- Funding. A new investment from Meta, announced Jun 12, 2025, after which Meta holds a minority of its equity.
- Public work. 20+ benchmark leaderboards and papers such as MCP-Atlas (Dec 18, 2025) and SWE Atlas (May 7, 2026).
- People. CEO Francis deSouza since Aug 10, 2026, with founder Alexandr Wang as chairman.
- Hiring. Open roles, among them Machine Learning Research Scientist, Post-Training.
Surge AI
Surge AI says its mission is to raise AGI with the richness of human intelligence. It sells human data and RL environments to AI labs.
- What it sells. RL environments and agents, rubrics and verifiers, RLHF and SFT data, human evaluation, multimodal data and off-the-shelf datasets.
- Who does the work. An expert workforce of 200,000+ PhDs across 500+ disciplines, as of Nov 19, 2026.
- Funding. Its site says it has never raised venture funding.
- Public work. Benchmarks such as DAYJOB for healthcare, finance, tax and law (September to November 2026) and EnterpriseBench Corecraft (Feb 18, 2026).
- Hiring. Open roles, among them RL Environments Architect.
Head-to-head comparisons
Our comparisons look at three of these companies side by side. See Mercor vs Scale AI, Mercor vs Surge AI and Surge AI vs Scale AI, or the alternatives guides for Mercor, Surge AI and Scale AI.
Comparison table
All twelve companies at a glance, in the order of this guide. Most recent dated news is the last dated post, round or deal on the company’s own site, as of Nov 19, 2026.
| Company | Group | What it sells | Most recent dated news | Public work |
|---|---|---|---|---|
| Patronus AITalk to us → | Research lab | RL environments, expert reviews and licensed data | Investigating What Open Models Knew When They Answered Questions Incorrectly, Nov 16, 2026 | FinanceBench, TRAIL, FigmaTrace, 20 papers (Nov 10, 2026) |
| Snorkel AI | Research lab | Datasets, benchmarks and evaluation environments | Paper2Agent: Turning Research Papers Into Interactive AI Agents, Nov 19, 2026 | Senior SWE-bench, Learning from Less |
| Datacurve | Coding specialist | Coding and data science data and environments | DeepSWE update, Sep 3, 2026 | DeepSWE |
| Handshake AI | Expert network | Expert datasets, evaluations and RL environments | Terminal-Bench 4 post-training result, Nov 19, 2026 | ATLAS benchmark suite |
| Mercor | Expert network | Expert datasets, benchmarks and RL environments | How frontier models miss risk in routine enterprise work, Nov 13, 2026 | APEX benchmarks, OWE-Bench |
| micro1 | Expert network | RL environments, evaluations and robotics data | Every task, checked against every public benchmark, Nov 18, 2026 | Dual-Arm Sim, MissionReport |
| Turing | Expert network | Datasets, RL environments and experts | Turn real company data into complete AI benchmarks, Nov 11, 2026 | Terminal-Bench Data, KernelQuest |
| Deccan AI | Full-service vendor | Training data, RL environments and evaluations | STARK RL Environments, May 25, 2026 | CASTBench, VoxClash TTS 2.0 |
| Invisible | Full-service vendor | RL environments, evaluations and annotation | How AI performance analytics is reshaping elite tennis at the ATP and Davis Cup Finals, Nov 19, 2026 | Language Atlas, data quality paper |
| Labelbox | Full-service vendor | RL environments and expert data | Inside Recursion Managed Agents: the architecture of reliable enterprise work, Nov 18, 2026 | EchoChain |
| Scale AI | Full-service vendor | Data, evaluations and RL environments | Measuring Robot Data Diversity, Nov 17, 2026 | Leaderboards, MCP-Atlas, SWE Atlas |
| Surge AI | Full-service vendor | Human data, rubrics and RL environments | DAYJOB: Law, Nov 18, 2026 | DAYJOB, EnterpriseBench Corecraft |
Where the market is heading
Funding rounds stayed large in 2026. Snorkel AI announced $350M on Sep 22, 2026, Labelbox more than $100M in October and Mercor a $750M Series D on Nov 10, 2026.
The vendors are also buying companies. Handshake acquired Cleanlab, Taro and Uplimit, Invisible agreed to acquire WeCP and Mercor acquired Deeptune, all in 2026.
Most of them publish benchmarks, such as Mercor’s APEX, Surge AI’s DAYJOB, Handshake AI’s ATLAS and Datacurve’s DeepSWE, so a buyer can see a sample of their work in the open.
How to choose an AI training data startup
The same questions work whether you are buying data or deciding where to work.
Ask who writes the data
Ask what share of tasks come from practicing professionals, how they are vetted and how many of them review each task.
Ask how quality is measured
Ask whether the vendor trains models on its own data and what the models gain on held-out tasks. Patronus AI publishes how it measures this.
Read the public work
A team’s benchmarks and papers let you see how it approaches problems before you meet it. Every entry above names its public work.
Check where the data comes from
Ask who owns the source material and on what terms it was licensed, and who owns the data from a custom build.
For candidates, ask what you would build
Ask which domains you would work on and whether the team publishes. Each entry above says whether it is hiring.
Where Patronus fits
Patronus AI sits in the research labs group. Its team publishes benchmarks such as FinanceBench, TRAIL and GlobeBench, and builds RL environments and the data inside them.
Most of its work is custom. It builds the tasks in each environment, calibrates them to what a team needs and usually sends samples first.
To talk about a project, contact Patronus AI. To join the team, see open roles.
Frequently asked questions
What is an AI training data startup?
A company that sells the data AI labs train and test models on, such as worked examples, ranked answers, evaluations and RL environments. Most work with networks of domain experts. Patronus AI, Mercor, Snorkel AI and Surge AI are examples.
Which AI training data startups should I watch?
This guide covers twelve in four groups, the research labs Patronus AI and Snorkel AI, the coding specialist Datacurve, the expert networks Handshake AI, Mercor, micro1 and Turing, and the full-service vendors Deccan AI, Invisible, Labelbox, Scale AI and Surge AI.
Do AI training data companies sell RL environments?
Most do. As of Nov 19, 2026, all twelve companies in this guide list RL environments on their own sites. Our guide to RL environment startups covers startups that build RL environments as their main product.
Which AI training data startups publish research?
Most of them. Patronus AI publishes FinanceBench, TRAIL and GlobeBench, Mercor the APEX benchmarks, Surge AI DAYJOB, Handshake AI the ATLAS suite, Datacurve DeepSWE and Snorkel AI Senior SWE-bench.
Which AI training data startups are hiring?
All twelve list open roles as of Nov 19, 2026. Patronus AI lists its roles on its Join page, and each other entry names an example role or the kinds of roles it hires for.
How do I compare Mercor, Scale AI and Surge AI?
Start with our head-to-head pages, Mercor vs Scale AI, Mercor vs Surge AI and Surge AI vs Scale AI.
How was this list made?
Patronus AI wrote it and is one of the companies on it. Entries are grouped by how they work, not ranked. Each company’s own site says it sells training data, evaluations or RL environments to AI labs, and each fact comes from the company’s own site or announcement, a paper it published, or the profile its entry names.
Snapshot as of