Each forgery in the source data patches a single field into an otherwise authentic scan - roughly 0.3% of the page. At a 224px whole-page input that edit survives as a handful of pixels, and a random-resized crop can miss it altogether. So every row here is a native-resolution window around the edit, paired with a second window from the same page, cut where the mask says no pixel was changed and matched on ink density so both crops carry comparable content.

That same-page negative is deliberate, not a fallback. It shares its positive's document, scanner, paper, font, JPEG history and encoder, so none of those can separate the classes - a separate authentic file would match on none of them for free. It is a different window, not the same one: the identical region of an authentic counterpart does not exist here, since v1's authentic tree cannot be joined to its own metadata and v2 ships none at all.

Crops are stored lossless and are never resized: resampling is what destroys the high-frequency artifacts the detector reads.

column	meaning
`image`	the crop, PNG bytes
`mask`	tampered pixels, 255 = tampered (all-255 for global-tier rows)
`label`	0 authentic, 1 tampered
`tamper_scope`	`local` (a field was edited) / `global` (the page is synthetic) / `none`
`pair_id`	the source document; both members of a pair share it
`split`	derived from the source document id, never from the upstream partition
`backend_model`	which generator produced the forgery
`tamper_bbox`	edited field in crop coordinates, `[x, y, w, h]`

Splits

split is re-derived from the upstream document id, so a document lands in one split across AIForge v1 and v2. The partitions shipped by those repos are not used: they were assigned independently per repo, which puts the same receipt in v1's training set and v2's testing set.

Generators

gemini-2.5-flash-image - 3,547 forgeries
gpt-image-2 - 2,980 forgeries
gpt-4o - 776 forgeries
ideogram-v2-edit - 422 forgeries

Training on these means training on API-generated forgeries, which normally carry provenance watermarks. That is checkable rather than assumed here: v1 and v2 tamper the same documents from the same specs with different generators, so training on one and evaluating on the other reads out whether the detector learned the watermark or the tampering. Cross-generator collapse to chance is the signal that it learned the watermark.

Composition

split	authentic	tampered
train	4,816	5,377
val	756	818
test	1,377	1,530

Licence and attribution

Derived from, and inheriting CC BY-NC-SA 4.0 from:

Non-commercial use only; derivatives must carry the same licence. The upstream corpora (CORD v2, WildReceipt, SROIE, XFUND) carry their own terms. Cite the Scam-AI papers if you publish results.

Downloads last month: 21

Total file size:

2.8 GB