Dataset Viewer
Auto-converted to Parquet Duplicate
file
stringlengths
19
39
pages
float64
1
120
⌀
pdf_version
float64
1.5
1.7
⌀
xref
stringclasses
2 values
open_password
stringclasses
1 value
page_sizes_pt
stringlengths
7
959
⌀
page_rotate
stringclasses
8 values
markers
stringlengths
16
3.85k
⌀
features
stringlengths
10
155
a_correct_merge_keeps
stringlengths
9
107
bytes
int64
0
60.9k
note
stringclasses
6 values
files/plain_a4_3p.pdf
3
1.7
table
null
595x842 595x842 595x842
0 0 0
[plain_a4_3p p1/3] [plain_a4_3p p2/3] [plain_a4_3p p3/3]
baseline: A4 portrait, classic xref table
all pages in order
2,112
null
files/letter_2p.pdf
2
1.7
table
null
612x792 612x792
0 0
[letter_2p p1/2] [letter_2p p2/2]
baseline: US Letter portrait
page size 612x792 (not resized to A4)
1,652
null
files/mixed_sizes_3p.pdf
3
1.7
table
null
842x595 420x595 612x1008
0 0 0
[mixed_sizes_3p p1/3] [mixed_sizes_3p p2/3] [mixed_sizes_3p p3/3]
A4 landscape (by MediaBox), A5, US Legal in one file
each page keeps its own size and orientation
2,132
null
files/rotated_3p.pdf
3
1.7
table
null
842x595 842x595 595x842
90 270 0
[rotated_3p p1/3] [rotated_3p p2/3] [rotated_3p p3/3]
page /Rotate 90; /Rotate 270 inherited through a nested page tree node; 0
every TOP arrow still points up (inherited Rotate must be copied onto the page)
2,286
null
files/bookmarks_6p.pdf
6
1.7
table
null
595x842 595x842 595x842 595x842 595x842 595x842
0 0 0 0 0 0
[bookmarks_6p p1/6] [bookmarks_6p p2/6] [bookmarks_6p p3/6] [bookmarks_6p p4/6] [bookmarks_6p p5/6] [bookmarks_6p p6/6]
outline: 6 bookmarks in 2 levels, explicit /XYZ destinations, opens with the outline panel
all 6 bookmarks, nested the same way, each jumping to its page in the merged file
4,396
null
files/links_a_4p.pdf
4
1.7
table
null
595x842 595x842 595x842 595x842
0 0 0 0
[links_a_4p p1/4] [links_a_4p p2/4] [links_a_4p p3/4] [links_a_4p p4/4]
page 1 has 3 link annotations to named destinations section-1..3 (same names in links_a_4p and links_b_4p)
links on each contents page still open that file's own sections (name clash must be resolved)
3,247
null
files/links_b_4p.pdf
4
1.7
table
null
595x842 595x842 595x842 595x842
0 0 0 0
[links_b_4p p1/4] [links_b_4p p2/4] [links_b_4p p3/4] [links_b_4p p4/4]
page 1 has 3 link annotations to named destinations section-1..3 (same names in links_a_4p and links_b_4p)
links on each contents page still open that file's own sections (name clash must be resolved)
3,236
null
files/page_labels_7p.pdf
7
1.7
table
null
595x842 595x842 595x842 595x842 595x842 595x842 595x842
0 0 0 0 0 0 0
[page_labels_7p p1/7] [page_labels_7p p2/7] [page_labels_7p p3/7] [page_labels_7p p4/7] [page_labels_7p p5/7] [page_labels_7p p6/7] [page_labels_7p p7/7]
page labels i, ii, iii, 1, 2, 3, A-1
labels of these 7 pages (offset by the pages before them)
4,019
labels: i ii iii 1 2 3 A-1
files/form_a_1p.pdf
1
1.7
table
null
595x842
0
[form_a_1p p1/1]
AcroForm with fields name, email, agree (checkbox), filled, with appearance streams; form_a_1p and form_b_1p use the same field names with different values
both forms show their own values after merging (fields renamed, or flattened)
2,874
values: Alice Example / alice@example.com / agree=True
files/form_b_1p.pdf
1
1.7
table
null
595x842
0
[form_b_1p p1/1]
AcroForm with fields name, email, agree (checkbox), filled, with appearance streams; form_a_1p and form_b_1p use the same field names with different values
both forms show their own values after merging (fields renamed, or flattened)
2,867
values: Bob Example / bob@example.com / agree=False
files/xref_stream_3p.pdf
3
1.5
stream
null
595x842 595x842 595x842
0 0 0
[xref_stream_3p p1/3] [xref_stream_3p p2/3] [xref_stream_3p p3/3]
PDF 1.5: cross-reference stream and an object stream (no classic xref table)
all pages
1,660
null
files/incremental_update_3p.pdf
3
1.7
table
null
595x842 595x842 595x842
0 0 0
[incremental_update_3p p1/3] [incremental_update_3p p2/3] [incremental_update_3p p3/3]
two revisions: revision 2 replaces page 1 (DRAFT -> FINAL) and appends page 3
the latest revision: page 1 says FINAL, 3 pages
2,866
revision 1 alone has 2 pages and says DRAFT
files/broken_xref_2p.pdf
2
1.7
table
null
595x842 595x842
0 0
[broken_xref_2p p1/2] [broken_xref_2p p2/2]
xref table offsets are all 7 bytes wrong (common after bad FTP/text-mode copies or editing)
readers that rebuild the xref open it; a merger should repair it or report a clear error
1,664
null
files/junk_before_header_2p.pdf
2
1.7
table
null
612x792 612x792
0 0
[junk_before_header_2p p1/2] [junk_before_header_2p p2/2]
80 bytes of mail headers before %PDF (offsets counted from %PDF)
both pages (Acrobat and most readers accept junk within the first 1024 bytes)
1,767
null
files/encrypted_no_open_password_2p.pdf
2
1.7
table
null
595x842 595x842
0 0
[encrypted_no_open_password_2p p1/2] [encrypted_no_open_password_2p p2/2]
RC4 128-bit, empty user password, owner password set, permissions deny print/modify/copy/assemble
both pages after decryption; the tool should ask for the password (second file) and must not output garbage
2,036
null
files/encrypted_password_test_2p.pdf
2
1.7
table
test
595x842 595x842
0 0
[encrypted_password_test_2p p1/2] [encrypted_password_test_2p p2/2]
RC4 128-bit, open password is: test
both pages after decryption; the tool should ask for the password (second file) and must not output garbage
2,020
null
files/shared_resources_120p.pdf
120
1.7
table
null
420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 420x595 ...
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[shared_resources_120p p1/120] [shared_resources_120p p2/120] [shared_resources_120p p3/120] [shared_resources_120p p4/120] [shared_resources_120p p5/120] [shared_resources_120p p6/120] [shared_resources_120p p7/120] [shared_resources_120p p8/120] [shared_resources_120p p9/120] [shared_resources_120p p10/120] [shared_r...
120 A5 pages that all share one Resources dictionary, one font and one logo XObject
one copy of the shared resources: merging this file twice should give about 2x the size, not 120x
60,928
null
files/contents_variants_3p.pdf
3
1.7
table
null
595x842 595x842 595x842
0 0 0
[contents_variants_3p p2/3] [contents_variants_3p p3/3]
page 1 has no /Contents (a valid blank page); page 2 /Contents is an array of 3 streams
3 pages, page 1 blank
2,483
page 1 is blank by design, so it has no marker text
files/metadata_unicode_2p.pdf
2
1.7
table
null
595x842 595x842
0 0
[metadata_unicode_2p p1/2] [metadata_unicode_2p p2/2]
Info and XMP title in UTF-16 (Cyrillic, Japanese, Spanish), Unicode bookmark titles
bookmark titles readable (not mojibake)
2,903
null
files/attachment_1p.pdf
1
1.7
table
null
595x842
0
[attachment_1p p1/1]
embedded file order.csv (document level attachment)
the page; ideally the attachment too (many mergers drop it, worth knowing)
1,623
null
files/annotations_2p.pdf
2
1.7
table
null
595x842 595x842
0 0
[annotations_2p p1/2] [annotations_2p p2/2]
page 1: sticky note, highlight (no appearance stream), external URI link
all 3 annotations on page 1
2,183
null
files/truncated_3p.pdf
null
null
null
null
null
null
null
plain_a4_3p.pdf cut at 2/3: no xref, no trailer, last object incomplete
a clear error message (or, for the truncated file, the pages that can be recovered); no crash
1,408
invalid input
files/zero_bytes.pdf
null
null
null
null
null
null
null
empty file
a clear error message (or, for the truncated file, the pages that can be recovered); no crash
0
invalid input
files/png_named_pdf.pdf
null
null
null
null
null
null
null
a 1x1 PNG with a .pdf extension
a clear error message (or, for the truncated file, the pages that can be recovered); no crash
70
invalid input

PDF Merge Test Files

24 small PDF files (130 KB in total) that each break PDF mergers in a different way, and 16 merge cases with the expected result. Every page prints a marker such as [bookmarks_6p p3/6], so after merging you can check page order with plain text extraction, or by eye from the coloured band at the top of each page.

The set covers rotated pages (including /Rotate inherited from a nested page tree node), mixed page sizes, bookmarks, links to named destinations that clash between two files, form fields with the same names in two files, page labels, PDF 1.5 cross-reference streams, an incremental update, a broken xref table, junk before the header, two encrypted files, 120 pages sharing one set of resources, pages with no content or split content, Unicode metadata, an attachment, annotations, and three files that are not valid PDFs at all.

All files were written byte by byte by a script. No third-party documents. CC0.

Why this dataset

Merging looks like "copy the pages one after another", and on clean files that works. Real files are not clean: they come from scanners, mail clients, form tools and old editors, and the bugs show up as missing bookmarks, links that jump to the wrong file, forms where both copies suddenly show the same name, upside down pages, or a 60 KB input turning into a 7 MB output. This set is useful for:

  • Testing a merge feature in a library, a web app or a command line tool. merge_cases.csv lists which files to merge, how many pages must come out and the markers in the expected order, plus what else to look at.
  • Regression tests. The files are tiny and deterministic, so they fit into a repository and a CI run.
  • Checking robustness of PDF parsers. The damaged and invalid files should give a clear error or a repaired result, never a crash or an empty output.
  • Seeing what gets lost. Bookmarks, page labels, attachments and form fields are often dropped silently. Each row of files.csv says what a correct merge keeps, so you can decide knowingly what your tool supports.

Layout

files.csv          one row per file
merge_cases.csv    16 suggested merges with the expected result
files/<name>.pdf   the test files
file what it tests
plain_a4_3p, letter_2p baselines: A4 and US Letter
mixed_sizes_3p A4 landscape, A5 and US Legal in one file
rotated_3p /Rotate 90 on a page, /Rotate 270 inherited from a nested Pages node, no rotation
bookmarks_6p 6 bookmarks in two levels, opens with the bookmark panel
links_a_4p, links_b_4p a contents page linking to named destinations section-1..3; both files use the same names
form_a_1p, form_b_1p filled form fields name, email, agree with the same names and different values
page_labels_7p page labels i, ii, iii, 1, 2, 3, A-1
xref_stream_3p PDF 1.5 cross-reference stream and object stream, no classic xref table
incremental_update_3p revision 2 replaces page 1 (DRAFT becomes FINAL) and appends page 3
broken_xref_2p every xref offset is 7 bytes wrong
junk_before_header_2p mail headers before %PDF
encrypted_no_open_password_2p RC4 128-bit, opens without a password, owner restrictions on print, copy, modify and assemble
encrypted_password_test_2p RC4 128-bit, open password test
shared_resources_120p 120 A5 pages sharing one Resources dictionary, font and logo
contents_variants_3p page 1 has no /Contents (a valid blank page), page 2 has an array of three content streams
metadata_unicode_2p title and author in Cyrillic, Japanese and Spanish in Info and XMP, Unicode bookmarks
attachment_1p a document level attachment order.csv
annotations_2p sticky note, highlight and an external link on page 1
truncated_3p, zero_bytes, png_named_pdf not valid PDFs: cut at two thirds, empty, a PNG with a .pdf extension

files.csv columns: pages, pdf_version, xref (table or stream), open_password, page_sizes_pt (as displayed, after rotation), page_rotate, markers (expected marker text per page), features, a_correct_merge_keeps, bytes, note.

merge_cases.csv columns: case, inputs (in order), expected_pages, expected_markers, what_to_check. Invalid files count as zero pages in expected_pages.

How the files were checked

  • pdf.js opened every valid file (with the password where needed) and found the right number of pages, the markers in order, the displayed page sizes and rotation, the latest revision of the incremental update, 6 bookmarks, the page labels, the form fields, the attachment, the three annotations and links that land on pages 2, 3 and 4. The password protected file refuses to open without the password.
  • pypdf 5.1 merged all 16 cases. Every result had the expected page count and the markers in the expected order (checked again with pdf.js). Merging shared_resources_120p twice gave 124 KB from a 61 KB input. pypdf rejected the three invalid files with an error, and opened broken_xref_2p and junk_before_header_2p only after repairing them.
  • Pages were rendered in Chrome and looked at.

What is not here

No AES-256 encryption, no XFA forms, no digital signatures, no linearized files, no tagged PDF structure trees and no optional content layers. Files are small on purpose, so this is not a performance benchmark.

Merging these files in a browser

To merge some of these files by hand, use https://toolster.net/merge_pdf : add the files, drag pages into order, rotate or remove pages, and save. The files are processed in the browser and are not uploaded.

Downloads last month
55