pdfredeval

An evaluation framework for third-party PDF redaction tools

redaction-tools.com · github.com/RedactionTools/pdf-redaction-benchmarks

Public web services, desktop software and vendor APIs, measured as black boxes: no source, no internals, only what a submitted page comes back looking like.

Part of Redaction Tools — results are published to the catalogue there.

Design, metric definitions and dataset structure live in docs/.

Install

pip install pdfredeval                    # CLI + Python API: generation and adapters
pip install "pdfredeval[score,report]"    # + scoring and HTML report charts

Extras: score, report, generate (rasterised conditions), datasets, publish.

Setup

For development, the project is managed with uv.

uv sync                    # create .venv, install the project editable + dev tools
uv sync --extra generate   # + Pillow, for the rasterised probe conditions
uv run pytest              # 300 tests
uv run ruff check .
uv run mypy

Scoring needs a PDF parser and a rasteriser — pypdf, pypdfium2, pandas, numpy — and the report's charts need matplotlib. The score and report extras declare them and the dev group pulls both in, so a plain uv sync can run the whole suite. Generation and the adapter layer still need nothing. The OCR leak layer shells out to tesseract if it is on PATH; without it that layer is reported unavailable rather than clean, because "found nothing" from a pass that never ran is the same sentence as "no leak".

uv run <cmd> syncs first, so a fresh clone needs no separate install step. uv.lock is committed: this project's whole premise is that a result names what produced it, and a benchmark whose own toolchain drifts cannot hold tools to that standard.

Try it

uv run pdfredeval families
uv run pdfredeval generate all --seed 1-3          # or one family: pii-detection
uv run pdfredeval inspect pii-detection-1 --probes

With no output flags, everything lands in one committed tree per package version (see Benchmark data). generate writes benchmarks/v<version>/cases/<family>/<case_id>/{<case_id>.pdf, ground_truth.json}, and stamps the version folder as the case's dataset revision. The same seed and timestamp reproduce byte-identical output on any machine — every page is stamped with its generation time, and ground_truth.json records it so --generated-at replays a case exactly. The full condition matrix (23 cells) needs the generate extra for its rasterised cells; without it those are reported on stderr as skipped, never silently omitted.

Driving a tool works best with a registered adapter, but none ship in-tree. For a tool without one, submit lays out a manual run: the folder, screenshots/, a TASK.md for the operator, and the handle.json that collect needs. It claims every channel and rendering condition, so no probe is excused as unsupported:

uv run pdfredeval tools
uv run pdfredeval submit acme:web pii-detection-1 --operator you
#   ... follow <run_dir>/TASK.md, drop the tool's PDF into <run_dir>, any name ...
uv run pdfredeval collect <run_dir>
uv run pdfredeval score                  # every run of this version, then its reports

score also takes a bare PDF with --case-dir, which needs no adapter at all — and scoring a case against itself is the do-nothing tool, so it must report a leak rate of 1:

uv run pdfredeval score some/dir/pii-detection-1.pdf \
    --case-dir some/dir

It writes score/{probes.csv, report.json}: one row per probe, and a report carrying the threshold table, the engine versions and the alignment it was computed with. Every rate comes with a Wilson interval and its n, because 0 leaks in 40 probes is [0, 0.088], not 100%.

score then reports on what it wrote: a summary in the terminal, report.md and a self-contained report.html — no CDN, no sibling charts/ directory, so it survives being forwarded — beside a single run; for a sweep of the tree, under its reports/ per tool, per tool and family, and across tools (--report-out, or ./report, for runs outside the tree). Each report also draws ground truth over the tool's own output (leaks boxed red, redactions green) unless you pass --no-overlay; --no-report stops after the scored artefacts:

uv run pdfredeval score <run_dir> --no-overlay       # numbers only, no images
uv run pdfredeval score some/runs/ --title "October sweep"     # pooled into ./report

A target that already carries score/ is reported from those artefacts, never re-scored — a published number must not move because tesseract or pypdf drifted underneath it. Pass --rescore to score such runs again; it asks first. Counts pool across runs; per-run rates are never averaged.

Publishing

Sign in once per machine. Your browser opens on redaction-tools.com; sign in there and approve the code your terminal shows:

uv run pdfredeval login                         # --no-browser over SSH: prints the link
uv run pdfredeval publish <run_dir> --dry-run   # check, list, send nothing
uv run pdfredeval publish <run_dir> --notes "Pro plan, default settings"
uv run pdfredeval publish-screenshots <run_dir>  # screenshots for a run already published
uv run pdfredeval logout                        # revokes this machine's key

The key is saved in ~/.config/pdfredeval/credentials.json (readable by you only; $PDFREDEVAL_CONFIG_DIR moves it). For CI, where there is no browser, create a key on your account page and set PDFREDEVAL_API_KEY=<prefix>.<secret> instead: the environment is read only, never a flag, and it wins over a saved login.

Runs are grouped into one submission per tool and dataset revision. Each run sends its manifest.json, score/report.json, the delivered PDF and its overlay, and nothing else: never ground_truth.json, probes.csv, screenshots or entities.json. The site rescores every PDF with the same scorer version and badges the result verified or disputed. An editor reviews it before it reaches the leaderboard. --site or $PDFREDEVAL_SITE points at another deployment, for example http://localhost:8007.

Benchmark data

Generated cases, tool runs, scores and reports are committed, one tree per package version, because a score is only comparable against the exact cases it was run on:

benchmarks/
  LATEST                                    newest version folder
  v0.1.1/
    README.md                               index: cases per family, runs per tool, reports
    cases/<family>/<case_id>/               <case_id>.pdf, ground_truth.json
    runs/<vendor-surface>/<run_id>/         manifest.json, the tool's PDF (any name),
                                            TASK.md, screenshots/, score/, report/
    reports/<vendor-surface>/               one tool, every family
    reports/<vendor-surface>/<family>/      one tool, one family
    reports/_summary/[<family>/]            every tool

Case PDFs are named by case id, so the family is in the name: a PDF uploaded to a vendor or attached to a ticket still says what it is. A run's output keeps whatever name the tool gave it - the run folder already names the tool - and collect records that name in the manifest. Keep one PDF per run folder; with two, collect asks which.

• --bench-root (or $PDFREDEVAL_HOME) moves the tree; --bench-version picks another version folder. -o, --runs-dir and --out still write anywhere else.
• Nothing already there is overwritten unasked. At a terminal you are asked to overwrite, skip or abort; elsewhere (CI, pipes) the command stops unless you pass --force or --skip-existing. Reports are derived from score/ and are simply redrawn.
• Trees and runs from before this layout still load: a case's case.pdf and a run's output.pdf are read under their old names.

Layout

Status

Implemented: case generation, the adapter layer, the aligner, the probers, the scorer, the reporter and the publisher.

Not yet: object probes (faces, signatures, QR codes, stamps), which leaves the image, attachment and font-subset leak layers implemented but with nothing to catch; and the raster-dependent rendering conditions (handwriting, photo grounds, degraded scans), which are reported as skipped rather than silently omitted.