LaTeX translation test files (BookTranslator, 2026-10-26). Free to reuse.

sample-source.tex              98-line English LaTeX file written for this test (Shelfmark is a
                               fictional app and the paper is invented): macros, an abstract,
                               labelled sections, an equation, display and inline math with \text{},
                               a footnote, \ref, \eqref and \cite with ties, an itemize list, a
                               booktabs table, a figure with caption, two comments, a verbatim block
                               containing a bare %, \url, \href, \$ and \&, and an inline bibliography
run_llm_latex_test.py          runs the three prompt conditions against a local OpenAI-compatible
                               server (whole-file, instructed, segments)
outputs/*.tex                  the nine Simplified Chinese outputs, unedited (moved into outputs/
                               after the run)
check_latex_translation.py     structural checker (Python standard library)
check-results.txt              checker output for all nine, run with --keep Shelfmark
compile_check.py               compiles each output as returned and again with
                               \usepackage[UTF8,fontset=fandol]{ctex}, then counts unresolved
                               references, unresolved citations and missing glyphs
compile-results.txt            compile_check.py output for all nine
results.csv                    one row per run, including details the checker does not report

Model: Qwen 3.6 27B (qwen/qwen3.6-27b, Q4_K_M) in LM Studio, reasoning_effort "none",
temperature 0.3, Apple M4 Max with 64 GB of memory. One request per run for whole-file and
instructed; one request per block for segments. Compiled with Tectonic 0.17.0 (XeTeX engine);
PDF text read with pdftotext (Poppler). The instructed and segments prompts tell the model not
to change the preamble; that rule is part of the result, not a bug we fixed afterwards.