prosperious's picture
Add 1000 files (OCR_INDRA/.gitignore ...)
f9bc0c1 verified
|
Raw History Blame Contribute Delete
2.88 kB

Limbu OCR research paper

The manuscript is DRAFT.md. It presents LimbuOCR as an operational, deployed page-to-book research platform — Sirijonga recognition, page/book assembly, crop-linked human review, the explicitly unresolved Devanagari-written Limbu evidence boundary — together with the gated Nepal language-pack pipeline built around it (per-script silver substrate, staged training payloads, one measured second-language diagnostic, and two root-caused negative results, each reported with its admission status), and the research program toward low-review-burden book digitization.

Supporting files:

  • ARTIFACTS.md maps each numerical statement to a surviving artifact and records its claim limitation.
  • LIMBU_SOURCE_VERIFICATION.md records hashes and bounded visual checks for the Sirijonga newspaper source and the CDC visual/reference mismatch.
  • build_heldout_page_subset.py selects the three page dates reserved by the recognizer split.
  • build_figures.py regenerates seven figures from local metrics: six Limbu figures plus the language-pack substrate/status chart.
  • verify_claims.py checks source counts, metrics, payload configuration, model hashes, runtime paths, source identities, paper values, and figure outputs.
  • AI_INTEGRITY_CHECK.md binds any detector-risk result to an exact corrected PDF hash. A receipt for an earlier manuscript is not valid for this one.

Build

make -C paper-buildkit

The build performs no OCR inference or training. It regenerates figures from existing JSON evidence, verifies the manuscript, and produces paper/draft.pdf with Pandoc, citeproc, and XeLaTeX.

To regenerate the three-date page manifest before rescoring existing document JSON:

python3 paper-buildkit/build_heldout_page_subset.py

The exact score commands are retained in each directory under paper-buildkit/evidence/. The score runs are non-oracle but light: they reuse exported page documents and do not invoke the recognizer.

Evidence status

The paper is a systems-and-research-platform draft, not a SOTA submission claim. The operational platform exists, while the Sirijonga results remain real-print, date-separated silver diagnostics. They need native review, a larger independent page set, external baselines, book-scale review-burden measurement, and resolved data rights. The bundled Devanagari runtime does not yet have a clean Limbu-specific benchmark on visibly Devanagari pages. The language-pack section reports substrate, staged payloads, and diagnostics only — every non-Limbu artifact sits behind an explicit admission gate (rights review, native label review, or both), and the section claims no per-language quality result.