neurotic_docx_bench: results and outputs

Run records and outputs of neurotic_docx_bench, a benchmark that scores DOCX tools against Microsoft Word: DOCX to PDF converters against Word's own PDF export, and redline engines against Word's own compare, with Word rendering both sides. Failures count as zero. The benchmark is maintained by Arthur Souza Rodrigues (Jandira Technologies), who also writes Jubarte, one of the ranked tools; Jubarte rows are marked † and follow the same rules.

Related repos

Repo	Role
arthrod/neurotic_docx_bench-fixtures	Inputs: source DOCX fixtures, mutation probes, and Microsoft Word oracle renders (PDF/DOCX)
superdoc-dev/docx-corpus	Upstream source of the real-world .docx documents (ODC-By)
arthrod/docx-corpus	Copy of the docx-corpus metadata under this account
arthrod/contractscrub-repro	Sibling legal-document benchmark (LLM contract review)
arthrod/sec-ex10-exhibits	Sibling legal corpus: SEC EX-10 material contracts as Markdown
GitHub: jandira-tech/neurotic_docx_bench	Benchmark code, RESULTS.md, methodology

What is here

Path	Contents
bench.jsonl	One line per recorded run: tool, version, document set, renderer, scorer and per-run statistics
outputs/redlines_0929_full/<tool>/	Each tool's redline DOCX for every compare pair, and the PDF Word made of it
site/	The cases behind jubarte.pro/cases: source documents, Word's oracle, each tool's output and scores (site/manifest.json, with a sha256 for every file). See site/README.md
LICENSE-ODC-BY-1.0.txt, NOTICE	Licence and attribution for the collection

Input fixtures and Word oracle renders are in arthrod/neurotic_docx_bench-fixtures.

Configs (viewer)

The original files are unchanged. The data/ parquet configs were derived from them (with tools/to_parquet.py) so the Hub viewer can show them; every config has a single train split.

Config	Rows	Built from	One row is
runs (default)	392	bench.jsonl	One recorded run: tool, version, benchmark, summary statistics, configs as JSON strings
run_doc_scores	146,044	bench.jsonl (scores, per_doc, timings)	One (run, document) score
redline_0929_pair_scores	19,123	outputs/redlines_0929_full/run/scores_*.json	One (tool, compare pair) in the 2026-09-29 full redline run, scored or missing
redline_0929_failures	2,425	outputs/redlines_0929_full/*/meta/generate_failures.json	One pair a tool failed to redline, with the error
site_cases	1,095	site/manifest.json	One case shown on jubarte.pro/cases (860 convert + 235 redline) with every engine's score
files	29,426	repo tree	One DOCX/PDF file in outputs/ or site/, with size, hash and download URL

Runs by benchmark: script_redlines 343, accepted_changes 18, roundtrip 12, visual_redlines 7, visual_rendering 7, visual_accepted_changes 5 (2026-07-08 to 2026-08-14). Vendors: jubarte-rust 155, jubarte 136, docxodus 28, jubarte-wasm 23, jubarte-ast 23, superdoc 9, folio 7, redlines 3, docx-redline-js 3, superdoc-redlines 2, sanity-word 2, ooxmlsdk 1.

Schema

runs

Column	Type	Description
id_run	string	UUIDv7 of the run
timestamp	string	ISO-8601 UTC
vendor, tool_version	string	Tool and pinned version/commit
benchmark	string	script_redlines, accepted_changes, roundtrip, visual_redlines, visual_rendering, visual_accepted_changes
n_docs, n_scored	int64	Documents in the set / documents with a score
overall_mean, overall_median, min, max, std, q1, q3	float64	Pixel score (0–100) statistics over scored documents
exact_100, at_least_90, below_50	int64	Counts by score band
page_mean, page_median	float64	Page-level score statistics
*_speed	float64	Throughput statistics recorded by the harness
itt_mean, itt_median, itt_n_docs, n_failures	float64/int64	Intent-to-treat statistics (failures = 0), newer runs only
scorer, skill_mean, skill_median, v2_mean, v2_median	mixed	Scorer version and alternative scores, newer runs only
n_functional_checked, n_accept_ok, n_reject_ok, n_lens_disagree, lens_disagree_rate	mixed	Functional accept/reject checks, newer runs only
n_oracle_unmatched, corpus_revision, holdout_mode	mixed	Corpus bookkeeping, newer runs only
config_hash, detail	string	Hash of the run config; path of the per-run detail file in the GitHub repo
failures_json, n_failures_listed	string, int64	Failure list as JSON (doc, stage, error) and its length
score_config_json, environment_config_json, build_recipe_json	string	Scorer parameters, render/generate commands, build recipe (JSON)

run_doc_scores

Column	Type	Description
id_run, vendor, benchmark, tool_version, timestamp	string	Joins to runs
doc	string	Document / pair stem
score	float64	Pixel score 0–100
page_count, overall_score_strict, overall_score_drift, min_score	int32/float64	From per_doc (newer runs only, else null)
generate_s, render_s, raster_s, score_s	float64	Stage timings in seconds (when recorded)
pages_json	string	Per-page detail as JSON (when recorded)

redline_0929_pair_scores

Column	Type	Description
source_file	string	The scores file the row came from
tool	string	jubarte-0.10.1, jubarte-rust, jubarte-093, docxodus, superdoc
variant	string	redline (compare), accepted / rejected (accept-all / reject-all of the redline)
subset	string	full or sample500_seed0
pair	string	<original>__vs__<modified>_redline_<hash>
status	string	scored, or missing (scored 0 under ITT)
overall_score, average_score, min_score (+ _strict, _drift, _pagefair variants)	float64	Pixel scores vs Word's compare, rendered by Word
ink_jaccard, text_boundary	float64	docxide-pdf's metrics (0–1)
page_count, page_count_candidate, page_count_oracle, page_count_mismatch, max_break_drift	int/float/bool	Pagination checks
raster_ns, score_ns	float64	Timings in nanoseconds
oracle	string	corpus or fresh
null_score, score_v2, skill_score	float64	Always null in this run

redline_0929_failures

Column	Type	Description
tool	string	superdoc (2,419) or docxodus (6)
pair	string	Compare pair stem
stage	string	Pipeline stage that failed
error	string	Error message

site_cases

Column	Type	Description
id	string	Case id (<state>-<hash> for convert, r-<hash> for redline)
bench	string	convert or redline
stem	string	Corpus stem (docx-corpus SHA-256 based)
state	string	clean, tracking_without_comments, with_comments_tracking, with_comments_clean
files	list	Repo paths of the case's files
engines	list<struct{engine, engine_label, score, failed, page_scores, jaccard, text_boundary}>	One entry per engine
extra_json	string	Any other manifest fields (currently {})

files

Column	Type	Description
path, filename, ext	string	Repo path, file name, docx/pdf
section	string	outputs/redlines_0929_full, site/convert, site/redline
tool	string	Tool folder under outputs/ (null under site/)
subdir	string	Sub-folder (docx, pdf_by_word, accepted/by_word, …) or site case id
size_bytes	int64	File size
sha256	string	From outputs/…/run/MANIFEST.sha256.json or site/manifest.json (null where not listed)
git_blob_id	string	Git blob id on the Hub
url	string	Direct download URL

Examples

runs (long JSON strings truncated):

{"id_run": "019f443e-58ca-7288-91d7-8acd3e50b9bf", "timestamp": "2026-07-09T00:19:24.490489+00:00", "vendor": "jubarte", "tool_version": "jubarte-final@717311c03d4f", "benchmark": "accepted_changes", "n_docs": 166, "n_scored": 166, "overall_mean": 78.1534, "overall_median": 80.639, "exact_100": 26, "at_least_90": 43, "below_50": 14, "min": 37.7791, "max": 100.0, "q1": 69.3773, "q3": 91.4018, "config_hash": "700d20b3cb79", "detail": "results/detail/019f443e-58ca-7288-91d7-8acd3e50b9bf__accepted_changes.json.gz", "failures_json": "[]", "score_config_json": "{\"align_upsample\": 10, \"blob_min_size\": 40, …}", "itt_mean": null, "scorer": null}

redline_0929_pair_scores:

{"source_file": "outputs/redlines_0929_full/run/scores_jubarte-0.10.1.json", "tool": "jubarte-0.10.1", "variant": "redline", "subset": "full", "pair": "0018843794_889f4c39…__vs__943e62959d_88646e92…_redline_49aa6b8d2f", "status": "scored", "overall_score": 50.06, "average_score": 54.68, "min_score": 39.30, "ink_jaccard": 0.431, "text_boundary": 0.030, "page_count": 11, "page_count_candidate": 11, "page_count_mismatch": false, "oracle": "corpus"}

redline_0929_failures (local path truncated):

{"tool": "docxodus", "pair": "19eb128df9_98bf5f3d…__vs__9666fa13fb_a3701d36…_redline_5b269ade28", "stage": "generate", "error": "docxodus-csharp-inproc: timeout waiting for worker reply after 120000 ms …"}
{"tool": "superdoc", "pair": "cfd53bad25_behavior__sd_2672_sdt_table_fc038913__vs__671ff91b1d_behavior__sd_2750_borderbox_7cd3768d_redline_001b877031", "stage": "generate", "error": "diff.apply: tracked shared definition replacement is not supported. No mutation was performed. …"}

site_cases:

{"id": "r-ecad91f16e", "bench": "redline", "stem": "00f467c010_0008753a…__vs__aa5340125d_00a4b0b9…_redline_ecad91f16e", "state": "tracking_without_comments", "files": ["site/redline/r-ecad91f16e/original.docx", "site/redline/r-ecad91f16e/modified.docx", "site/redline/r-ecad91f16e/word.docx", "site/redline/r-ecad91f16e/word.pdf", "site/redline/r-ecad91f16e/jubarte.docx", "site/redline/r-ecad91f16e/jubarte.pdf", "…"], "engines": [{"engine": "jubarte", "engine_label": "jubarte 0.10.1 (release)", "score": 91.17, "failed": false, "jaccard": 95.13, "text_boundary": 100.0}, {"engine": "docxodus", "engine_label": "Docxodus 12.6.5", "score": 72.81, "failed": false, "jaccard": 82.69, "text_boundary": 82.63}, {"engine": "superdoc", "engine_label": "superdoc-sdk 2.16.0", "score": 0.0, "failed": true, "jaccard": null, "text_boundary": null}]}

files:

{"path": "site/convert/clean-010cd893de/docxide.pdf", "section": "site/convert", "tool": null, "subdir": "clean-010cd893de", "filename": "docxide.pdf", "ext": "pdf", "size_bytes": 75550, "sha256": "bd217ab051172293dee68e9dc2cf0cf3add47ce60214437fd2ed54e503e385d0", "url": "https://huggingface.co/datasets/arthrod/neurotic_docx_bench/resolve/main/site/convert/clean-010cd893de/docxide.pdf"}

Usage

from datasets import load_dataset
from huggingface_hub import hf_hub_download

runs = load_dataset("arthrod/neurotic_docx_bench", "runs", split="train").to_polars()
cases = load_dataset("arthrod/neurotic_docx_bench", "site_cases", split="train")

# fetch one case's files
case = cases[0]
paths = [hf_hub_download("arthrod/neurotic_docx_bench", f, repo_type="dataset") for f in case["files"]]

Headline results

From RESULTS.md generated 2026-10-01 22:11 UTC. Compare rows only within one table; full history, paired comparisons and methodology are in RESULTS_DETAILED.md.

Every corpus DOCX to PDF vs Word's own PDF (6,427 documents; ITT = failed documents score 0)

Rank	Tool	Pin	Run	Docs	Failed	ITT Mean	ITT Median	95% CI	Mean	Median	Perfect (100)
1	jubarte †	jubarte 0.10.1	2026-10-01	6427	0	77.03	80.90	[80.58, 81.19]	77.03	80.90	38
2	soffice	LibreOffice 26.8.0.3 bce0998afefdbc355585ca324285661a2170ba77	2026-10-01	6359	68	61.95	63.43	[62.84, 64.03]	62.61	63.72	4
3	docxide-pdf	docxide-pdf v0.17.1	2026-10-01	6422	5	54.14	51.44	[51.01, 51.84]	54.18	51.45	16

Redlines vs Word's compare (3,502 pairs; in parentheses, only the pairs Docxodus redlined)

Tool	Version	Scored	ITT Mean	ITT Median	= 100	>= 90	< 50	Ink Jaccard mean	Text boundary mean
jubarte-0.10.1	jubarte 0.10.1 (release)	3496/3502 (2610/2612)	74.82 (75.48)	84.34 (84.61)	241 (174)	636 (472)	617 (435)	0.73 (0.74)	0.80 (0.81)
jubarte-rust	jubarte 0.10.0 (86b6b5d3)	3499/3502 (2611/2612)	72.23 (73.30)	81.73 (83.01)	310 (244)	658 (517)	793 (558)	0.67 (0.69)	0.77 (0.78)
jubarte-093	jubarte 0.9.3 (release)	3338/3502 (2584/2612)	68.62 (72.44)	79.07 (82.61)	309 (254)	634 (524)	780 (562)	0.67 (0.68)	0.76 (0.78)
docxodus	Docxodus 12.6.5 (C#)	2612/3502 (2612/2612)	53.94 (72.33)	62.87 (81.75)	204 (204)	482 (482)	614 (614)	0.65 (0.65)	0.73 (0.73)
superdoc	superdoc-sdk 2.16.0	380/3502 (323/2612)	6.66 (7.60)	0.00 (0.00)	4 (0)	7 (3)	20 (18)	0.21 (0.21)	0.29 (0.27)

Limitations

• The score is pixel similarity to Microsoft Word's output. It has not been validated against human judgment of redline quality; a correct redline can lay out differently from Word's compare.
• The oracle is one build of Word for Mac; Word-rendered rows can only be reproduced on a machine with Word.
• Adapters for other tools are written by the Jubarte author. Known adapter faults, including ones first found by a tool's maintainer, are recorded in docs/VENDOR_NOTES.md.
• Failures count as zero; both intent-to-treat and scored-only columns are printed.
• The sealed holdout has never been run.
• In redline_0929_pair_scores, missing rows are pairs with no scorable output; the headline tables count them as 0. superdoc's 3,122 missing pairs are mostly generate failures listed in redline_0929_failures.
• runs mixes harness generations: columns such as itt_*, scorer, skill_* exist only for newer runs.
• Some JSON fields (environment_config_json, failure messages) contain the author's local paths; they are build metadata, not data.

Licence and source documents

Contains information from docx-corpus (superdoc-dev/docx-corpus, built by SuperDoc), made available under the ODC Attribution License (ODC-By v1.0). ODC-By covers the collection, its selection and arrangement; it does not license copyright in the individual .docx files, which were collected from the public web and may carry their own rights. A rights holder who wants a document removed can write to contact@arthur.law; it is removed from this dataset, the fixtures and future runs. The benchmark code is AGPL-3.0.

Links

• Code and generated results: https://github.com/jandira-tech/neurotic_docx_bench
• Side-by-side viewer: https://jandira-tech.github.io/neurotic_docx_bench/
• Fixtures: https://huggingface.co/datasets/arthrod/neurotic_docx_bench-fixtures