Datasetpublic

midah/ai-data-deals

AI Training Data Deals A small, curated dataset of public AI training data licensing deals — buyers (model developers), sellers (content providers), use case (training / RAG / fine-tuning), pricing mechanism, exclusivity, modality, financial terms when disclosed, and per-field provenance. Provenance Every deal carries the source URL it came from and the exact passage (source_text) that justified the fields. Where financial terms are disclosed, the financial_terms… See the full description on the dataset page: https://huggingface.co/datasets/midah/ai-data-deals.

sourceHugging Facecc-by-4.0updated 5mo agoView on Hugging Face

0likes12downloads

Dataset Card

AI Training Data Deals

A small, curated dataset of public AI training data licensing deals — buyers (model developers), sellers (content providers), use case (training / RAG / fine-tuning), pricing mechanism, exclusivity, modality, financial terms when disclosed, and per-field provenance.

Provenance

Every deal carries the source URL it came from and the exact passage (source_text) that justified the fields. Where financial terms are disclosed, the financial_terms struct preserves the original passage and a confidence level (high, medium, low).

When a field is unset, missing_reasons records why — undisclosed, not_applicable, unknown, or potential_future (for deals that were announced as planned rather than executed). This lets analyses distinguish "this field is missing because we haven't extracted it yet" from "this field is missing because the deal is private."

Sources

Records are aggregated from four sources, deduplicated by canonicalised buyer + seller. The confirmed_by column lists every source that confirmed each deal.

—ithaka — SR Ithaka generative-AI licensing tracker
—monda — Monda's list of AI data licensing deals
—neudata — Neudata AI data licensing market analysis
—seed_cross_referenced — manual entries cross-referenced against primary press coverage

Schema

Column	Type	Notes
`buyer`	string	Canonicalised model developer (e.g. `"OpenAI"`, `"Google"`)
`seller`	string	Content provider
`date`	string	`YYYY`, `YYYY-MM`, or `null`
`use_case`	string	`training`, `rag`, `fine_tuning`, `combined`, `unknown`
`pricing_mechanism`	string	`per_unit`, `aggregate`, `service`, `commissioning`, `open_commons`, `unknown`
`exclusivity`	string	`exclusive`, `non_exclusive`, `unknown`
`modality`	string	`text`, `image`, `audio`, `video`, mixed combinations
`financial_terms`	struct or null	`{value, source_url, source_text, confidence}`
`creator_compensated`	bool or null	Whether original creators (not just intermediaries) are paid
`missing_reasons`	list[struct]	Each entry is `{field, reason}`
`source_url`, `source_text`	string	Provenance for the deal record itself
`ingested_from`	string	Pipeline source
`fetch_date`	string	When the row was last refreshed
`confirmed_by`	list[string]	Sources that independently confirmed this deal

Limitations

—Coverage is not exhaustive — only public deals are tracked, and only those reported by one of the four sources above.
—Financial terms are reported by the source. They are often imprecise ("undisclosed, multiyear"), estimated by trade analysts, or derived from earnings filings.
—Buyer/seller canonicalisation is rule-based, not exhaustive.