Marker 2.0.0: lesson test 2026-10-14 (passed)

Cite: headlesson test:48hvric — Marker 2.0.0, 2026-10-14: 4/4 lesson steps passed (operator record). On 2026-10-14, the headlesson operator ran all 4 steps of this lesson with Marker 2.0.0 (macOS 26.5.2 arm64; Python 3.11.13 isolated venv; pinned wheel-only installation; preloaded official font and models; CPU with two owned localhost model servers; one authored English digital PDF; default autospawn not tested; no optional remote LLM processing), and all passed. The final Next step is advice and was not run. No independent check has been reviewed. This is the operator's own run, not a vendor claim. Sample tested: one self-made English digital PDF page (three text blocks and a 3-row by 2-column table). https://headlesson.com/tests/48hvric

Record test:48hvric. Question: Convert an owner-chosen local PDF without enabling optional remote LLM processing.

Answer

On 2026-10-14, the headlesson operator ran all 4 steps of this lesson with Marker 2.0.0 (macOS 26.5.2 arm64; Python 3.11.13 isolated venv; pinned wheel-only installation; preloaded official font and models; CPU with two owned localhost model servers; one authored English digital PDF; default autospawn not tested; no optional remote LLM processing), and all passed. The final Next step is advice and was not run. No independent check has been reviewed. This is the operator's own run, not a vendor claim.

Tested on

  • Test date: 2026-10-14
  • Record confirmed: 2026-10-14
  • Version: 2.0.0
  • Environment: macOS 26.5.2 arm64; Python 3.11.13 isolated venv; pinned wheel-only installation; preloaded official font and models; CPU with two owned localhost model servers; one authored English digital PDF; default autospawn not tested; no optional remote LLM processing
  • Sample tested: one self-made English digital PDF page (three text blocks and a 3-row by 2-column table)
  • Result: passed, 4/4 steps

Lesson steps and what happened

  • 1. [passed] pip install marker-pdf
  • 2. [passed] marker_single --help
  • 3. [passed] marker_single ./sample.pdf
  • 4. [passed] Inspect the generated output against the source PDF.

Next (advice, not tested)

Next: Select the real PDF and destination, run marker_single <confirmed-PDF-path>, and compare text, tables and equations with the original. Choose --output_format json for structured output; only enable --use_llm after agreeing on provider, data transfer, cost and model-weight terms.

Cautions recorded with the lesson

These notes come from the lesson. Some describe failures seen in tests; others are general cautions. Each one is not a separately tested result.

  • Python 3.10 or newer, PyTorch and an owner-provided local sample are prerequisites.
  • Apache-2.0 and free code use do not establish unconditional rights to model weights. Modified model-weight terms and paid commercial licensing may apply.
  • The optional --use_llm path can incur provider costs; the initial conversion steps do not enable it. Next treats LLM use as optional and requires the user's data-transfer, cost and model-term decision.
  • The bundled REST server supports fewer flags than CLI/SDK and omits --use_llm and --disable_ocr.
  • Conversion output is not independent evidence of transcription accuracy.
  • The 2.0.0 trial needed official font, layout and OCR-error model assets even for a digital PDF. These were preloaded and two owned localhost model servers were used through supported settings; automatic download and autospawn were not tested. The source code's default batch-service autospawn can keep a server running and write shared home-cache logs, so account for cleanup.
  • Confirm: Check output location, recognition quality on this document, provider credentials/costs and applicable model-weight terms; optional LLM use is not an accuracy guarantee.

What this record does not show

  • Not measured in this run: Japanese text.
  • Not measured in this run: scanned pages.
  • Not measured in this run: math.
  • Not measured in this run: complex tables.
  • Not measured in this run: real user documents.
  • Not measured in this run: the automatic font and model download and the default server start (autospawn).
  • Not measured in this run: the optional --use_llm path.
  • Whether an agent new to this tool can use it from this record alone (the operator wrote and ran the lesson).
  • Comparisons between models or clients.
  • Production use, accounts, or real customer data.
  • Behaviour of the latest version (the test is tied to the version and date below).
  • The final "Next" step is advice and was not part of the test.

Primary record

Operator record: a past lesson trial (legacy). It is not claimed to meet level L1 of run-task-contract/0.1.0.

Independent check

Independent check: not yet reviewed. (Whether an independent check exists has not been looked up for this record.)

Evidence

A non-public test report exists and is not published, so third parties cannot check it directly.

Retrieve over MCP

headlesson_get id="app:marker" sections ["lesson","gotchas"] on https://mcp.headlesson.com/mcp

Source version

This page is built from the catalog record app:marker at revision f9af4e5f7493. It is not rewritten when the current lesson changes; a new test gets a new record ID.

Sources (lesson text at the fixed revision)

Each line names fields of this page (as JSON pointers into the .json version) and where they come from. Labels without a link are non-public sources. Fields not listed (such as test results) are the headlesson operator's own observations.

Corrections

None.

Page updated: 2026-10-15 (the date its content last changed).

要約(日本語): Marker の lesson を 2026-10-14 に headlesson の運営が 2.0.0 で試し、4/4 段が通った(passed、運営の実測で製品側の主張ではない)。この記録が示さないこと・独立確認の状況は本文のとおり。