Where AI training data comes from, provider by provider.

“Every task is written by PhD researchers, practicing lawyers, and senior engineers, not crowdsourced.” source

“Training data, evals, and RL environments built by the same experts behind our frontier-lab work.” source

“Expert data and evaluation environments built with leading domain specialists across 100+ fields” source

“Domain experts curate artifacts that capture the complexity, ambiguity, and edge cases of real professional work.” source

“Verified domain experts generating SFT data, instruction-response pairs, and specialist annotations across 80+ languages.” source

“We deliver expert data development as a service: datasets, benchmarks, evaluations, and environments built with our experts, rather than a SaaS app your team runs.” source

“Image, video, text, document, speech, and sensor data labeled by trained subject-matter experts across 120+ languages” source

“Research-grade and built by experts to hillclimb on key benchmarks.” source

“TaskUs is training its teams of subject matter experts to accelerate training data delivery using V7’s model-assisted labeling stack.” source

“Expert preference rankings, comparative evaluations, and nuanced feedback from verified subject matter experts across medicine, law, science, and finance for reinforcement learning with human feedback.” source

“to curate problems that experts outside of Proximal are uniquely aware of.” source

“We source elite domain experts for on-demand projects.” source

“We work with domain experts to capture that thinking” source

“Premium multimodal data from a global network of creative professionals.” source

“Now, Toloka collaborates with data workers from 100+ countries speaking 40+ languages across, 50+ knowledge domains and 120+ subdomains.” source

“Currently, our model for producing these tasks involves each task being owned by a single engineer, who is responsible for all phases of the task lifecycle” source

“Developed in collaboration with domain experts, our datasets are carefully curated to be of extremely high quality” source

“Powering the world’s leading AI companies, governments, and research institutions with data sourced from verified contributors.” source

“It was built through a globally distributed workforce of tens of thousands and a continuous, standardized collection process that keeps capture conventions consistent across the dataset.” source

“Vetted domain experts matched to your task, across physical AI, multimodal systems, and LLMs.” source

“We build each environment with practitioners from the job, who ground the case, the documents and the checks in how the work is really done.” source

“We run your policies across our global robot fleet, with expert operators trained in the exact tasks being run.” source

“A fully managed workforce experienced in using client-developed and all major 2-D and 3-D tools.” source

“Tap into a crowd of 1.6M+ global experts spanning 150+ countries and 500+ languages, dialects, and locales” source

“We power RLHF pipelines at scale by combining expert raters, multilingual communities, safety frameworks, and proprietary orchestration.” source

“A collection of expert conversations across a range of domains.” source

“Physicians, attorneys, engineers and scientists help train frontier models with us” source

“Domain experts interested in building RL environments and data pipelines.” source

“15,000+ red teamers generating continuous adversarial data” source

“a global, vetted contributor network that produces datasets to a buyer's spec” source

“Every scenario written by a domain SME from workflows they run daily.” source

“Custom-sourced global subject experts for multimodal supervised fine-tuning, RLHF, red teaming, alignment and validation.” source

“Datacurve provides expert quality code data at scale from highly skilled software engineers.” source

“Tasks in MAGI are written by expert operators who spend two to three days immersed in the full Company Soul environment before writing a single prompt.” source

“Two multimodal reviewers scanned the sheets independently without seeing the target finding and nominated candidate slices.” source

“we work with domain experts to understand these discrepancies and construct detailed, information-dense environments.” source

“Our evaluators are domain experts who’ve done the work your agent is trying to imitate.” source

“we’ve signed national level partnerships to scale our contributor network to 50,000+ people.” source

“atomic action annotations with full expert validation” source

“Our first 30M cell atlas represents thousands of hours of human labor - the combined efforts of the Pythia Biosciences, Miraomics and LatchBio curation teams.” source

“They’re experts in their subject matter, with highly advanced reasoning, analysis and judgement skills.” source

“Every company, record and rule inside is authored by domain experts” source

“Exclusive, global network of over 500 domain experts” source

“Get responses for hard problems across domains (coding, STEM, agentic use cases, etc) curated by experts.” source

“Physicians, attorneys, and investment analysts with years of high-stakes experience, supported by a platform built for structured, high-quality data collection at scale.” source

“Domain specialists authored all tasks and verified every rubric criterion.” source

“We then leverage human expert audio annotators and our own annotation platform to process audio data for Automatic Speech Recognition.” source

“Expert humans label what machines miss: intent, context, edge cases.” source

“This is expert input, not crowd labor.” source

“We're hiring full-time Robot Operators to power our real-world data operations.” source

“We collect this data through more than 300,000 active contributors who record conversations, everyday tasks, and other project-specific activities through our private apps.” source

“Finance environments need people who have actually lent, plus researchers and engineers to build them.” source

“The initial release contains expert-authored evaluation items for Europe and India, with goal-specific rubrics and no model responses or attack trajectories.” source

“Collaborate closely with environment and research teams, frontier labs, and Fleet’s expert network to translate research needs and model failures into shared system capabilities.” source

“Run expert sessions, capture trajectories, verify outputs, and adjudicate edge cases.” source

“we’re building a community of the world’s brightest minds to create the next generation of training data.” source

“Our skilled data annotators meticulously create high-quality training datasets from various data types, including text, images, videos, LIDAR point cloud, and more” source

“Your training data is produced by vetted domain experts with on-chain reputation, not anonymous crowd workers.” source

“Long-horizon RL environments for code, computer-use and enterprise workflows that challenge SOTA models, built on a network of 300,000+ domain experts.” source

“Our in-house Subject Matter Experts (SMEs) help create ground truth for your agentic data.” source

“Our expert team labels actions, objects, and environmental context with hierarchical accuracy checks.” source

“Our managed global crowd deliver accurate datasets for even the most complex AI use cases.” source

“We work with a vetted network of expert contributors to collect high-quality computer-use data across” source

“the programs through which experts contribute to our datasets” source

“Model reasoning traces are evaluated against ground-truth game states using expert-curated rubrics and variance-checked grader panels” source

“We source domain experts, verify them through live proctored testing and interviews, then manage your team day-to-day inside your tools.” source

“We directed knowledge-work professionals to refine those rubrics around real-world priorities.” source

“Every datapoint is created from scratch by senior practitioners, primarily senior software engineers alongside vetted domain specialists.” source

“Every scenario is independently authored by a domain expert, passes through two levels of expert review and editorial compilation, and gets Pass@8 sampled to confirm adversarial validity and quality.” source

“Refresh builds environments by partnering with expert engineers in our network to source real tasks that break the leading models.” source

“Today, those same engineers author the datasets, RL environments, and evals that leading AI labs use to train and evaluate their models.” source

“Domain-matched US professionals delivering context-aware, high-accuracy labeled data.” source

“We’ve built a network of 20k+ experts across STEM and professional services” source

“Shaip delivers RLHF preference data, rewrites and rationales from vetted subject-matter experts, calibrated, quality-audited and formatted for your training pipeline.” source

“Instead of crowdsourcing your data needs to anyone and hoping for the best, we deliver AI training data collected, annotated, and validated by our TrainAI community of 100,000+ active, vetted, skilled, and qualified AI data specialists based on your specific ML project requirements.” source

“Tap into our global network of contributors and specialized experts to generate supervised fine-tuning (SFT) data and execute reinforcement learning and human feedback (RLHF) workflows.” source

“We are a team of PhD-level mathematicians and engineers tackling research-level open problems.” source

“Data across many languages, built by native experts who encode each one's grammar, idiom, and worldview, not translations.” source

“Domain experts and native speakers are selected from our 500K+ vetted workforce.” source

“Demonstrations, preference pairs and eval sets written by people who do the work every day.” source

“Real-world training data for robotics, VLA models, world models, wearables, and embodied AI.” source

“Participants recorded their spaces in 1080p from a first-person point of view to ensure real-world perspective.” source

“Custom data collection, annotation & evaluation” source

“Captured from domain experts step by step, decision by decision, and verified before delivery.” source

“Capture environment and task-specific data for your use case in labs and at scale.” source

“Toloka's self-serve platform gives you access to the same data collection infrastructure used by frontier AI labs” source

“All collected through distributed human-in-the-loop capture.” source

“Lightwheel has built a Physical Measurement Factory designed to systematically measure how real objects behave.” source

“Three pillars: data collection (in-field, in-lab, teleoperation, and synthetic capture), annotation with built-in QA, and evaluation services covering model output review, red-teaming, preference ranking, and benchmarking.” source

“delivering the highest-quality teleoperation, on-embodiment, and egocentric data” source

“Secure, ethical data collection services for speech, audio, image, video and multimodal data” source

“Data collection is guided by how models are expected to operate in production, capturing the inputs, contexts, and variability models will encounter in practice.” source

“Launch a targeted data collection.” source

“Luel operates a worldwide contributor network, hundreds of thousands of people across almost 100 countries, who are paid to record, upload, and license slices of their everyday lives.” source

“Our trained teams collect diverse, real-world first-person video tailored to AI training and evaluation needs.” source

“Full traces of expert execution, including tool calls, checks, pivots, and recoveries, so agents learn the shape of real work, not just the finished answer.” source

“2 million hours of first-person video of real manual labor, with IMU, 3D hand pose and frame-level action labels.” source

“The data was collected through Pango, a crowdsourced platform where users are compensated for contributing their natural computer interactions during actual work sessions.” source

“We started with our proprietary dataset of 186 de-identified CT, MRI, and PET/CT cases paired with radiologist reports.” source

“WhisperMind captures the way real hands move through real spaces, picking up a wet glass, opening a stuck drawer, finding the salt in a stranger's cupboard.” source

“The Service lets adult contributors record first-person video of everyday tasks and household activities, submit those recordings to us, and earn payouts for submissions that we approve.” source

“Clickworkers are registered freelancers who complete microtasks via our platform using an app or desktop.” source

“Through these partnerships, the company captures, structures, and annotates real-world industrial video from active production environments.” source

“Physical AI learns from physical labor, and we own every step of collecting the data” source

“We activate exclusive supply networks to capture and deliver high-quality, model-aligned content at scale.” source

“Recorded human gameplay with synchronized signals, so models learn how the world looks and how actions move through it.” source

“The capture app records demonstration sessions.” source

“A multimodal collection spanning text, images, video, LiDAR, and IoT sensor data” source

“First-person RGB video recorded during assigned tasks.” source

“We source long-form, natural conversations between real speakers so voice models learn interruptions, turn-taking, accent variation, and messy production speech.” source

“A global network of trained data collectors capturing real-world video across 6 continents.” source

“Contra Labs transforms real creative workflows from our network into AI training infrastructure.” source

“real-workplace human video with hand/body pose, depth, and subtask labels.” source

“Only IR videos and images were captured.” source

“Contributors capture egocentric task video, human manipulation, or robot teleoperation to the specification.” source

“Roll out an existing model on a physical robot.” source

“Our ethical scraping infrastructure is built for petabyte- and exabyte-scale procurements.” source

“Models and experts work in controlled environments while complete trajectories are recorded.” source

“Verified field teams capture multi-sensor data in the real world.” source

“The datasets and judgement they produce make AI work in their own languages.” source

“We can source or generate unique data in your domain and even film new datasets from scratch with our ‘Hollywood for AI’ approach.” source

“First-person (head-mounted or wearable POV) video capture of humans performing real tasks” source

“Multimodal data capture: audio, image, video, text” source

“and turn it into data models can learn from.” source

“the infrastructure captures interaction data to score models across two primary axes in our first evaluated game, Poker” source

“Contact-rich interaction data collected with proprietary tactile hardware.” source

“We provide AI training data that is collected, annotated, and validated by our active, vetted, and skilled community of AI data specialists, tailored to your specific machine learning project requirements.” source

“We capture and aggregate multimodal data across real-world, digital, and simulated environments.” source

“augmented with real-world experimental data and contextual” source

“Our extensive global network of contributors and specialized experts gather real-world data, including point-of-interest (POI), audio, and video, through remote and targeted field tasks.” source

“Reliable data from diverse sources, collected with automated and manual quality checks.” source

“Scripted and spontaneous recordings from diverse speaker populations.” source

“Engineers scripted synthetic logs (500 to 1,000 lines of normal operation mixed with 10-20 lines of bug symptoms)” source

“Context engineering, calibrated LLM-as-a-judge deployment, synthetic data generation and custom benchmarking for teams extending their research capacity.” source

“We use synthetic data in many different parts of our data generation and want to scale our methods to generate entire codebases from scratch.” source

“We curate expert-generated datasets, generate synthetic data, and help turn messy enterprise data into a form models can be trained on.” source

“Generate synthetic data from scratch, or validate your pre-generated data at any stage.” source

“Data is derived from or simulated after real-world sources” source

“Lightwheel builds SimReady assets, synthetic and egocentric human data, and evaluation platforms for Embodied AI.” source

“Three pillars: data collection (in-field, in-lab, teleoperation, and synthetic capture), annotation with built-in QA, and evaluation services covering model output review, red-teaming, preference ranking, and benchmarking.” source

“Generative Simulators extend this idea. Tasks, world dynamics and rewards are generated together, so difficulty can rise as the agent improves.” source

“Synthetically generated task instruction for training purposes. Represents the context of the full task.” source

“Synthetic and real human data at scale, including continuous human feedback from live games.” source

“Each is generated from a knowledge graph that serves as the single source of truth.” source

“Cua has 4K+ environments and 400K+ tasks spanning desktop apps, cross-app workflows, browser surfaces, and generated state variations.” source

“GaaS builds synthetic transaction environments to test fraud detection models against evolving scam tactics.” source

“Build reusable and reliable systems for producing high-quality synthetic data, tasks, and verifiers” source

“We built a synthetic data-generation pipeline that transformed production traffic into constrained, evaluable queries” source

“We generate reasoning traces for all of our tasks using DeepSeek-R1-0528.” source

“Scaling simulation engines in general is human capital intensive, using a weakly supervised approach we can artificially generate and validate hundreds and thousands of environments on demand.” source

“Fill data gaps and accelerate coverage with synthetic pipelines calibrated to your model.” source

“We bring data generation tooling, human experts, synthetic data augmentation, and rigorous quality control into one platform” source

“These are synthetic or task-derived environments.” source

“High-quality labeled datasets with real-world data and synthetic samples for AI models, machine learning, and generative AI” source

“We use coding agents as world-builders that construct environments, tasks, and their verifiers” source

“We license it, package it, and pay you for it.” source

“Enterprise sourced data from a wide-range of enterprises, covering different environments, industries and time-horizons” source

“Upload your work once and earn commission whenever it is licensed.” source

“Kled provides premium, unbiased data from top rightsholders to train AI models.” source

“We license enterprise data across industries and business functions.” source

“third-party listings, where external labs post their own datasets on Luel's marketplace for a revenue share” source

“Share a brief and we source data to your exact spec from our licensor network.” source

“We constantly look for new data partners with hard-to-access or in-demand data to add to the millions of patient journeys we already have.” source

“We license private codebases and company operating data to train and evaluate AI models.” source

“not synthetic, not scraped.” source

“When we license code or data from a third party to build a benchmark, that agreement governs what we receive.” source

“Ooak Data licenses the operational and business data that defines how companies actually work, anonymizes it, and turns it into training material for frontier AI models.” source

“Petrarch sources suppliers, cleans and restructures the datasets, certifies legal clearance, and resells to applied AI startups and labs.” source

“Earn revenue by licensing content you already own, or by recording the skilled work your team already does.” source

“Enterprises come to build reliable, domain-specific AI grounded in their own data and operating knowledge.” source

“help turn messy enterprise data into a form models can be trained on” source

“whether sourced from the marketplace or uploaded from their own systems, ensuring data is model-ready before it enters training pipelines.” source

“Founders working with SimpleClosure can recover it through two distinct ways: the codebases they built, and the workspace data from how their teams actually worked.” source

“we present a collection of 200 environments drawn from the company's operational data over the past decade” source

“Troveo matches you with them, anonymizes your data to a documented standard, and licenses it on terms you approve” source

“We dig into your gameplay data, find what is genuinely useful to AI labs, and help you license and monetize it.” source

“Leverage Anthromind experts to convert your data into business value assuring your competitive edge.” source

“Turn the operational data organizations already hold into models, evals, and decision tools that do their actual work.” source

“We train on your data and deploy on your own infrastructure, so your data and the models built on it stay yours.” source

“We package your agent's interaction data for AI labs.” source

“We synthesize your first-party data into custom RL environments that reflect real goals, constraints, and outcomes closing the gap between simulation and production.” source

“Symbal partners with enterprises, converting your first-party data into a live digital map of your organization's processes.” source

“datasets that customers provide for annotation, evaluation, red-teaming, or related services” source

“Expert-built tasks run inside high-fidelity enterprise app clones” source

“They mirror real computer and API-based systems, supporting rich logs and application state that can be used for programmatic and rubric-based evaluation.” source

“Interactive replicas of enterprise and consumer apps such as Jira, Salesforce, and Zendesk.” source

“We've built 100s of gyms that mimic popular software like Slack and” source

“MAGI consists of 1,000 (200 test / 800 train) tasks mimicking real work spread across 8 cloned enterprise applications.” source

“Platform engineers build replicas of platforms, populated by data from the knowledge graph.” source

“Stateful sandboxes / clones of enterprise tools with state persistence across turns.” source

“Environments that replicate your target deployment” source

“High-fidelity clones of full software suites, wiring together multiple connected applications, from EHRs to enterprise apps and beyond.” source

“we built realistic, high-fidelity simulated replicas of email and calendar web applications” source

“Sandboxed replicas of the tools agents use at work: CRMs, ticket queues, spreadsheets, codebases.” source

“Key datasets and foundational research will remain publicly available.” source

“Kaggle OULAD (Open University Learning Analytics Dataset) for Learning Management System data, Kaggle HR Analytics for enterprise HR data” source

“The first batch of EgoSuite-Open100K is live on the Hugging Face Hub, open for academic research and commercial training, with the remainder rolling out in stages.” source

“The data used for these evaluations is from publicly available manuscripts, case studies, or databases.” source

“Each dataset contains 1000 validated tasks harvested from popular GitHub repos.” source

“Our data comes from our own lending and from public records, and we own all of it.” source

“By aggregating data from publicly available datasets, using existing research repositories,” source