Toolster/pdf-signing-test-documents

PDF Signing Test Documents

17 PDF documents with visible "Sign here", "Date" and "Initials" boxes, built to catch the bugs that make a signature land in the wrong place: rotated pages, MediaBox that does not start at 0,0, CropBox smaller than the page, rotation inherited from the page tree, a content stream that leaves a scaled coordinate system behind, mixed page sizes, a form with an empty signature field and an image-only scanned page. For each of the 40 boxes, boxes.csv gives the position both as the user sees it on screen and in PDF coordinates, and signed/ holds a reference copy of every document with the boxes filled correctly.

There are also 80 signature image files: 12 synthetic signatures and 4 sets of initials, each as transparent PNG, PNG on white, JPEG on white, a JPEG that looks like a photo of paper, and SVG.

Everything was generated by a script. The documents are fictional (a school trip form, a room rental agreement, a freelance contract, an invoice approval). The signatures are procedural pen strokes, not anyone's real handwriting, and they do not spell real names. CC0.

Why this dataset

Placing an image on a PDF page sounds simple until the page is rotated. The click happens in screen coordinates (origin top left, y down, after rotation and cropping), but the PDF wants user space coordinates (origin bottom left, y up, before rotation). Getting that conversion wrong is the most common bug in signing tools, and it only shows on certain files. This set is useful for:

• Testing a "sign PDF" or "stamp PDF" feature. Put a signature into each box using box_x_pt, box_y_pt, box_w_pt, box_h_pt (screen coordinates in points) and compare your output with signed/. If the signature ends up sideways, upside down, half size or outside the page, the case name tells you which conversion is missing.
• Unit tests for coordinate conversion. Every row has both systems: screen points and relative values (0..1) on one side, pdf_x0 pdf_y0 pdf_x1 pdf_y1 in PDF user space on the other, plus the page's MediaBox, CropBox and /Rotate. The values were checked with pdf.js (convertToViewportRectangle) for all 80 placements, unsigned and signed.
• Signature field detection. 16_acroform_signature_field.pdf has an empty /Sig field and a text field at the boxes, so you can check whether a tool finds the field instead of asking the user where to sign.
• Signature image import. The same signature comes with alpha, on white, as JPEG (no alpha, so the white must be removed) and on paper with a ruled line, uneven light and a slight tilt. signatures.csv has the ink bounding box of each clean image, which is the correct crop.

Layout

boxes.csv                     one row per box (40 rows)
signatures.csv                one row per signature image file (80 rows)
documents/<case>.pdf          unsigned documents
signed/<case>.pdf             the same documents with every box filled correctly
signatures/<variant>/<name>   png_transparent, png_white, jpeg_white, jpeg_scanned, svg

document	what makes it tricky
01_letter_portrait, 02_a4_portrait, 03_legal_portrait	nothing, these are the baselines (US Letter, A4, US Legal)
04_a4_landscape_mediabox	landscape because the MediaBox is wide, /Rotate 0
05_a4_rotate90, 06_a4_rotate180, 07_a4_rotate270	portrait page stored with /Rotate 90, 180, 270
08_offset_mediabox	MediaBox [200 300 795 1142], so (0,0) is outside the page
09_cropbox_inside_a3	A3 MediaBox, only an A4 CropBox is visible
10_cropbox_rotate90	offset CropBox and /Rotate 90 together
11_inherited_rotate	/Rotate and MediaBox sit on the Pages node, the page itself has neither
12_unbalanced_cm	the page content sets a 0.5 scale and never restores it; an appended stamp not wrapped in q/Q comes out half size
13_mixed_pages	4 pages: A4, Letter landscape, A5 rotated 90, A4; initials on every page, two signatures on the last
14_two_signers	two signature boxes and two date boxes on one page
15_long_document	24 pages, the only boxes are on the last page
16_acroform_signature_field	empty signature form field and a Date text field
17_scanned_form	a 150 dpi grey scan tilted 0.8 degrees, no text layer; the box coordinates are the bounds of the tilted box

boxes.csv

column	meaning
document, page	file and 1-based page number
role, label	signature, date or initials, and the text printed in the box
page_rotate, mediabox, cropbox	the page geometry, with inherited values resolved
display_width_pt, display_height_pt	size of the page as shown, after CropBox and rotation
box_x_pt, box_y_pt, box_w_pt, box_h_pt	box on screen: origin top left, y down, in points (1/72 inch)
box_x_rel ... box_h_rel	the same divided by the displayed page size
pdf_x0, pdf_y0, pdf_x1, pdf_y1	the box in PDF user space, ready for an annotation /Rect
form_field	name of the form field at this box, if any
case, signed_reference	what the document tests, and the reference file

To convert screen points to pixels, multiply by your zoom (at 100 % one point is 96/72 pixels in most viewers).

How the files were checked

• pdf.js opened all 34 PDFs, and for every box the PDF rectangle converted to screen coordinates matches box_*_pt within 0.01 pt, the page size matches, the box label is found by text extraction and is upright on screen, and the form fields are found by name.
• pypdf 5.1 opens all 34 PDFs in strict mode without warnings.
• The signed references were rendered in Chrome and looked at, including the rotated, offset and scaled cases.

What is not here

No cryptographic (certificate based) signatures: this is about where a visible signature goes, not about PAdES. No encrypted files and no XFA forms. The signatures are synthetic scribbles, so they are not suitable for training signature verification or forgery detection models.

Signing these files in a browser

To try a document by hand, open it in https://toolster.net/sign_pdf : draw with the mouse or type a name, then save. The PDF is processed in the browser and is not uploaded.