AI training datasets from real companies

621 datasets on Grepped, from businesses that ran on the data for years: email, CRM, ERP, code, tickets, CAD, calls, video and more. AI labs license them for post-training, evals and RL environments.

620 of them show an indicative price. Prices are for one non-exclusive licence and are indicative until the owner signs off. Who each owner is, the full descriptions and the samples are shared with qualified buyers under our mutual NDA in Grepped Buyside.

By category

Company operations archive 316

Whole operating records of real businesses: email, documents, CRM, ERP, tickets, accounting and team chat, often over many years. AI labs use them to build RL environments and evals for agents that do real office work, and for post-training on long, messy, multi-system tasks.

Code & software 109

Private codebases with their history: repositories, pull requests and reviews, issues, tickets and product docs from real software companies. Used for coding agents, code review models, SWE-style RL environments and evals that public GitHub data can't provide.

Agent traces 35

Recorded sequences of people and software agents working through real tasks: browser and app sessions, tool calls and outcomes. Used to train and evaluate computer-use and tool-using agents.

Industrial & sensor 28

Data from factories, plants, engineering firms and connected devices: sensor streams, CAD and drawings, maintenance and manufacturing records. Used for industrial copilots, engineering agents and physical-world models.

Text & documents 24

Large bodies of professional writing and documents: publications, reports, manuals, contracts and expert material. Used for pre-training, post-training and domain evals.

Financial & transactional 22

Transaction records, financial models, ledgers, underwriting and accounting workflows from real finance businesses. Used for finance agents, spreadsheet and accounting evals, and RL environments built on real numbers.

Physical-world video 19

Video of real work in the physical world, including egocentric footage from workers, drones and vehicles. Used for robotics, world models and video understanding.

Expert workflows 15

How credentialed experts actually do their work, step by step: tasks, reasoning, worked answers and graded outputs. Used for expert-level evals and RL environments with verifiable rewards.

Consumer behaviour 14

How people search, browse, shop and engage, from consented panels, platforms and businesses. Used for recommendation, personalisation and consumer agents.

Speech & conversation 11

Recorded conversations and calls with transcripts and outcomes, across languages and industries. Used for speech models, voice agents and conversational evals.

Geospatial & imagery 6

Aerial, drone, satellite and mapped imagery with survey and job records. Used for geospatial models and visual inspection.

Healthcare & life sciences 6

Clinical, diagnostic and life-science data. Health data is handled case by case before anything moves to a buyer.

Video & imagery 6

Image and video collections, many with labels or edit events. Used for vision models, generation and evaluation.

Human preference / evaluation 5

Human judgements and preference pairs from real people at scale. Used for RLHF, reward models and evals.

Robotics 3

Robot teleoperation, manipulation and contact data. Used for robot learning and embodied AI.

Workforce & labour market 2

Longitudinal records of companies, roles and hiring. Used for labour-market models and HR agents.

By sector

Other ways in

Want this kind of data?

Qualified buyers see full listings, samples and quotes in Grepped Buyside. Or email e@grepped.ai.

Request access