EpubCCConverting a paper full of equations to EPUB
Updated:
Equations are the weak point: the EpubCC engine keeps the prose and simple inline symbols, but display equations with fractions, sums and subscripts come out as scrambled plain text, because it does not build MathML yet. On the DPO paper (arXiv, CC BY 4.0) our 8-gram recall is 0.814, behind OpenDataLoader (0.860) and PDFium (0.838).
Original and result
Display equations with fractions, sums, exponents and subscripts; Greek letters in running text; theorem-style derivations in the appendix.
Download source PDF Download the sample EPUB
When to keep the PDF
If the equations are the point of the document, keep the PDF for reading them. The EPUB is useful for the prose around them (introduction, discussion, related work), which reflows and joins across pages correctly.
MathML output is not planned for the first version; when it exists this case will be regenerated and the date above will change.
Source and licence
- Document
- Direct Preference Optimization paper (arXiv 2305.18290)
- Licence
- CC BY 4.0
- Pages
- 27
- Traits
- text PDF, one column, formulas, footnotes, academic paper
Rafailov, Sharma, Mitchell, Ermon, Manning and Finn, “Direct Preference Optimization: Your Language Model is Secretly a Reward Model”, arXiv:2305.18290v3, licensed CC BY 4.0 (checked on the arXiv abstract page, 2026-11-04). Pages shown unmodified; the EPUB is our conversion.
Settings used
- Engine: EpubCC engine at commit 3b74b26, the binary that generated the images and the sample (2026-11-04).
- Pages: the whole document for the figures; pages 4–5 for the images and the sample EPUB.
- Benchmark figures: a run of 2026-11-04 with the same binary, every engine on the whole document; this document is not yet part of the published benchmark.
- Options: none. The engine has no per-document settings yet; the same defaults apply to every file.
Validation and results
EPUBCheck EPUBCheck on our EPUB: 0 errors, 0 warnings.
EPUBCheck 5.4.0 on the downloadable sample: 0 errors, 0 warnings (checked 2026-11-04).
- Pages
- 27
- Time
- 0.15 s
- Peak memory
- 63 MB
- Chapters
- 11, from the paper's headings
- Footnotes found
- 7, 6 linked
- Equations converted to MathML
- 0
- Quality warnings
- none
Benchmark run of 2026-11-04, not yet part of the published benchmark.
| Engine | 8-gram recall |
|---|---|
| EpubCC engine (EPUB output) | 0.814 |
| EpubCC engine (plain text) | 0.810 |
| pdf_oxide · extract_text | 0.780 |
| pdf_oxide · to_plain_text (XY-cut) | 0.777 |
| pdf_oxide · Markdown | 0.807 |
| PDFium · text().all() | 0.838 |
| OpenDataLoader PDF (local mode) | 0.860 |
| pdftotext (Poppler) | 0.827 |
Limitations we observed
- Display equations are written as plain text in the order the PDF draws them: fraction bars disappear, numerators and denominators land on separate lines, and the equation number drifts («r(x, y) = β log + β log Z(x). (5) π ref (y | x)»).
- Subscripts are written inline with a space (π ref, π θ) instead of as subscripts; some superscripts are kept (π^SFT).
- No MathML: a reading app cannot render or read the formulas aloud.
- OpenDataLoader (0.860), PDFium (0.838) and pdftotext (0.827) have a higher 8-gram recall than our engine (0.814); the reference skips equations, so this measures the prose.
- The sample's title came out as «4 Direct Preference Optimization» (the first heading of the pages converted), not the paper's title.