Datasets:

---

docling-project

/

regression-dataset-for-docling-parse

Like 2

Follow

Docling 1.13k

docling-parse PR #327 fixes character decoding for base-14 font dictionaries
that carry the obsolete /Name entry (<< /BaseFont /ZapfDingbats /Name /F3 >>,
as ReportLab emits): the built-in encoding table was only consulted under the
font's /Name label, so ZapfDingbats code 108 (a71, the black-circle bullet)
fell back to StandardEncoding and decoded as l.

annots_01.pdf's /F3 is exactly this pattern, so the stored groundtruth
captures the misdecode: four list-bullet cells read l where the correct
decode is ● (U+25CF). This PR regenerates the four parser artifacts for that
page with the fix applied (via the --update-groundtruth write path).

Diff relative to main:

• .char.txt / .line.txt / .word.txt: the four /F3 cells change l -> ●;
  geometry untouched.
• .json.gz: text/orig of the same four char cells, four word cells and four
  textline cells; bitmap_resources[1].image.uri also re-encodes (verified
  pixel-identical, just non-deterministic PNG compression on rewrite).

Companion to docling-parse PR #327; once merged, that PR bumps
HF_DATASET_REVISION to the new revision.