Status at a glance¶
This page is for project managers and coordinators. It says what each tool is ready for, what has been measured, what is still experimental and what it depends on. It is updated with each release; every fact links to the page that holds it.
The six tools¶
All six are beta releases. A release's images are scanned before they are published and listed by digest (the release gate). Its API contract is checked against the previous release's, as each service's runbook describes under "OpenAPI".
| Tool | Pilot release | What has been measured | Experimental or not evaluated | Needs at run time | Limits |
|---|---|---|---|---|---|
| page-classification | 1.12.1-beta |
94.6–95.9 % of pages typed right first, on about 4,800 unseen pages per model | born-digital pages; the optional YOLO backend; the ensemble's own accuracy | nothing outside (models downloaded at first start) | README |
| ocr-postprocess | 1.15.1-beta |
rule changes, on 2,067 hand-labelled lines and over both collections | a full validation of the five labels; lines not in Czech | nothing outside (models downloaded at first start) | README |
| digital-convert | 1.4.1-beta |
tried on 15 AMČR excavation reports; no accuracy figure | the optional Docling engine; uploads over 250 MB | nothing outside | README |
| nlp-enrich | 1.3.1-beta |
nothing yet on archival documents | languages other than Czech | LINDAT's UDPipe and NameTag, or copies of them | README |
| keyword-extract | 1.6.1-beta |
nothing yet | the vocabulary-based keywords: a research output | a language-model service, for vocabulary-based keywords only | README |
| translator | 1.7.1-beta |
a small backend comparison, without reference translations | translation quality; the alternative backends | LINDAT's translation service, or a model server you run | README |
What each tool does, in plain words, and what a person should check: For archives.
The pilot¶
| When | What | Who |
|---|---|---|
| by 11 November 2026 | the six service images chosen and approved by digest | AMČR, ÚFAL |
| second half of November 2026 | acceptance on AMČR's own cluster, one result row per service | AMČR, ÚFAL |
| December 2026 – January 2027 | the S2 pilot: AMČR runs the six services itself | AMČR |
The handover, the runbooks and the contacts: S2 pilot handover.
What the tools depend on¶
- LINDAT's public services (UDPipe, NameTag, translation) have no service-level agreement. A deployment can run its own copies instead.
- The Hugging Face Hub supplies the models of page-classification, ocr-postprocess and keyword-extract when a service first starts. A deployment without internet access supplies them in advance.
- AMČR decides the rights in the outputs it publishes; the tools record their components and do not assert a licence for a result (Licences and paradata).
Open decisions and deferred work¶
- Per-document permission to send text to an external service: agreed with AMČR, not yet in the tools (Licences and paradata).
- The terms of LINDAT's output for the translator's default backend are still open (#78).
- The document-understanding benchmark, which compares off-the-shelf models with these tools, is deferred until after the pilot (#22).
Sources¶
| Source | What was taken from it |
|---|---|
the six tools' README.md, "Limits and maturity", at the releases in the table |
what is measured, what is experimental, what each needs |
the six tools' service/README.md, "Models and caches" |
what each service downloads and calls |
| S2 pilot handover | the releases, the dates, the approval by digest |
Operations; each service/README.md, "OpenAPI" |
the release gate; the API contract check |
| Licences and paradata | rights, and the planned per-document permission |
| hub issues #22, #78 | the deferred benchmark; LINDAT's output terms |