AI training data types
AI training data comes in distinct types, and each type is bought, priced and quality-checked differently. This index defines 13 of them and links to the providers that sell each.
Last reviewed
All data types
| Data type | Belongs to | Main buyers | How it is priced |
|---|---|---|---|
| RLHF and preference data | post-training data | Frontier labs and any team that post-trains a model | Priced per comparison or per hour of reviewer time. Expert or long-form comparisons cost far more than short crowd judgments. |
| Supervised fine-tuning data | post-training data | AI labs, enterprises fine-tuning open models | Priced per example, scaled by length, domain difficulty and reviewer seniority. |
| Expert demonstration data | post-training data | Frontier labs building agents and professional-grade assistants | Priced mainly by expert hourly rate, which tracks the profession: lawyers, physicians and quantitative analysts cost more than general contributors. |
| Reinforcement learning environments | post-training asset | Frontier labs training agents on software, browsing and office tasks | Typically priced per environment or per project, with premiums for realism, verifier quality and difficulty. |
| Synthetic training data | training data | Teams that lack real data in a domain or need privacy-safe substitutes | Often priced like software or compute rather than labor. Value rests on the generator, the filters and the proof that it improves a model. |
| Code training data | training data | Labs and companies building coding assistants and agents | Licensed corpora are priced per dataset. Expert-written solutions and reviews are priced per hour or per task. |
| Medical and clinical data | regulated training data | Healthtech companies and labs building medical models | Priced per dataset or access agreement, with cost driven by de-identification, consent, curation and clinician time. |
| Legal training data | professional-domain training data | Legal-tech companies and labs improving professional reasoning | Public-law corpora are cheap or free. Lawyer-created work product is priced at expert hourly rates. |
| Video training data | multimodal training data | Labs building video generation and understanding models | Commonly priced per hour of footage, with rates driven by exclusivity, resolution, rights clearance and uniqueness. |
| Speech and audio data | multimodal training data | Speech AI developers and voice-agent builders | Priced per hour of audio, with premiums for rare languages, accents, noisy environments and conversational dialogue. |
| Licensed text corpora | pretraining data | Foundation-model developers and search and answer products | Large publishers and platforms negotiate annual fees, some reported in the tens of millions of dollars. Smaller owners use marketplaces with revenue-share or per-use terms. |
| Multimodal training data | training data | Labs building vision, voice and omni models | Priced per item or per hour of media, with higher rates for dense human annotation. |
| Evaluation sets | evaluation asset | Labs, enterprises and auditors that need trustworthy model measurement | Priced per set or per project. Private, unreleased sets are worth more because public ones leak into training data. |
How these data types are priced
Pricing depends on whether the data is labor, license or software. Read how AI training data is priced for the full model, or see the auction mechanisms that can set the price.