Know what your collection data is worth to AI. Record by record.
Like a careful registrar, shap.li corrects your catalogue export, measures what every record contributes and prices the whole for AI buyers, before anyone else sees it.
Try the demo No sign-up. Synthetic data.
| Object number | Category | Maker | Date made | Donor email | Contribution (pts) |
|---|---|---|---|---|---|
| TM.1832.0422 | Steam engines | n/ablank | July 18321832-07 | empty | not measured yet+0.05 |
| TM.1805.0348 | Scientific instruments | Unknown | 09/09/18051805-09-09 | empty | not measured yet+0.18 |
| TM.1897.0043 | Photography | n/ablank | June 18971897-06 | lindqvist2@example.org | not measured yet+0.11 |
| TM.1808.0144 | STEAM ENGINESSteam engines | Whitcombe Instruments | 1808 | empty | not measured yet−0.07Check: neighbours say “Scientific instruments” |
| TM.1808.0144 | Steam engines | Whitcombe Instruments | 1808 | empty | Duplicate: removed |
| TM.1933.0398 | Railways | J. Harrison & Sons | 12/12/19331933-12-12 | empty | not measured yet+0.15 |
| TM.1927.0325 | PHOTOGRAPHYPhotography | N/Ablank | 1927 | empty | not measured yet+0.13 |
Three questions your register cannot answer on its own
What is wrong with it?
Upload an export as it is: Excel, CSV, JSON or Parquet. shap.li lists every problem in plain English, most serious first, with examples from your own file: personal data, duplicates that differ only by a space or a capital, placeholders typed for blanks, the same category spelt four ways, dates in four formats.
Safe fixes are ticked for you; anything that removes information is your call. Gaps are reported, never filled in. You get the cleaned file back in its own format, with a report of every change.
What does each record add?
Pick something a buyer might want to predict, such as an object’s category. shap.li measures each record’s share of how well the collection teaches it: its Shapley value, the fair-division method named after Lloyd Shapley, Nobel laureate in economics. The records that carry the signal rise to the top.
Records that make predictions worse are listed for checking, with the label their nearest neighbours carry. Usually that is a mislabel worth fixing before anyone pays for it.
What would buyers pay?
An indicative price for each kind of use (retrieval, fine-tuning, pre-training, synthetic data) from what the data measurably is and what you tell us about rights, rarity and expert annotation. Then a panel of simulated AI buyers, and the price that earns the most.
Simulated figures are marked as simulated, everywhere. Real deals recalibrate the model.
Nothing leaves without your say
- You approve every licence
- Buyers ask; you read who they are and what they want it for, then approve, decline or bargain.
- Delivered files can be traced
- Data files delivered to a licensee carry hidden marker records unique to that licence, so a leaked file can be traced back to it.
- Personal data is checked for before listing
- Contact details and other personal data are looked for in every sample, and a dataset flagged as holding them cannot be published.
- Files you clean are not kept
- Exports uploaded for cleaning are deleted after seven days, or the moment you ask. Only the record of what changed remains.
For AI teams
Licensable collections, with rights stated by the holder and a standard datasheet: what is in them, how complete, how varied, what you may use them for. Quotes are instant and graduated by volume; the licence ID travels with every file you receive.
Instant quote Synthetic example
| Collection catalogue, fine-tuning, 1 year | |
|---|---|
| First 50,000 records at 100% | £14,408 |
| Next 150,000 at 80% | £34,579 |
| 200,000 records, annual licence | £48,987 |
See it on a sample export
Play the museum: clean a messy catalogue, value its records, price it, approve a request. Or the AI lab: read a datasheet and get a quote. It takes five minutes.