github-head/docs/BODY_DEV_REVIEW.md

Historical body-text development review

This step prepares annotation work, not another model-quality score. It uses paragraph regions (including verse) from the pinned validation split of PiotrSty/impact-psnc-polish-ocr, excluding the three previous heading-control pages. Choose up to four regions per collection, round-robin across pages, in stable ID order. Selection does not depend on predictions or detector success.

python -m training.prepare_body_dev_review --output data/body-dev-review-v1

Requires Pillow, numpy and OpenCV; no GPU or model weights. Original JPG crops, source metadata and SHA-256 hashes are retained. PNG conversion preserves RGB pixels; line crops use the original RGB image without deskew or binarization. OpenCV normalization is used only to propose geometry, not alter saved pixels. Proposals preserve the full region width and divide vertical gaps at their midpoints; the first and last crop extend to the region edges. This avoids tight ink-band clipping but can retain neighbouring fragments, requiring visual review.

Review

1. Open region-review/index.html to inspect original paragraphs and references.
2. Open line-review/index.html to inspect proposed individual line crops.
3. Check the entire line is present, no adjacent text is included, and the reference really belongs to that crop. Preserve historical spelling.
4. Mark doubtful geometry or glyphs as requiring review. Do not silently replace private-use symbols or illegible text with a model prediction.
5. Export the JSON review history. Use independent reviewers/profiles before treating agreed labels as a candidate reference set.

The existing panel calls records pages; here each record is a region or line. All labels begin unverified. The line/reference alignment is a count-only proposal, not a certified correspondence. Models are not run or shown to reviewers. Count mismatches and overlapping boxes are logged in report.json, with their source regions still available for review; no reference strings are assigned to these line proposals. Do not discard these failures from a future full-page score.

Boundaries

candidate-lines.jsonl is NOT an evaluation or training release. Every record has eligible_for_evaluation: false; the generic review tool does not turn that flag on. Review/adjudication and a separate audited release step are required. Do not feed this file straight into an evaluator that ignores those fields.

Frozen test collection IDs and exact crop hashes are excluded. Near-duplicates, shared editions and upstream training overlap are not ruled out. In particular, similar collection names must not be assumed to identify independent works. These are historical paragraphs/verse, not a representative sample of modern Polish documents. Check typeface coverage manually before expanding or training.

Raw images, draft references and review histories remain local. Only preparation code, tests and methodology belong in the repository until publication is approved.

Local preparation, 2026-09-21

The data/body-dev-review-v2 run selected 19 paragraph regions across 10 pages and five collection IDs, with 96 reference lines. Eleven regions produced 63 count-matched line proposals across four collection IDs. Eight regions had line count mismatches, including every selected SLAWNA_VICTORIA_FT region. Therefore the 63 proposals are not representative of the entire selected sample; manual geometry work is needed before evaluating it as a whole.

The initial v1 run used tight detector boxes. Visual spot checks showed clipped letter tops; v2 uses full-width, inter-band midpoint crops. The original v1 artifacts were retained, not overwritten. Two v2 crops were visually spot-checked; neighbouring fragments and uncertain line boundaries still require human review. No reference was corrected and no review decision was fabricated.

Validation: 25 focused tests passed; all 200 output checksums verified; all 63 saved line images matched the pixels of the corresponding region crop exactly. Runtime: OpenCV 5.0.0, numpy 2.5.3. No model inference or training was performed. The existing review assets were reused unchanged; browser UI validation of this generated packet was blocked by the browser tool's local-file security policy.

Transcriptions entered as notes

Only after the reviewer explicitly confirms that notes contain replacement transcriptions, prepare a separate draft with:

python -m training.migrate_review_notes --manifest data/body-dev-review-v2/candidate-lines.jsonl --review reviewer-export.json --output data/body-dev-corrected-draft-v1 --notes-are-transcriptions

This preserves note text verbatim, the original manifest/export, prior decisions and per-record provenance. Blank notes leave the latest actual text edit intact. No verified events are synthesized: all new records remain drafts, ineligible for evaluation. The new review panel starts a separate confirmation history; old statuses remain in source_review_decision and original-review.json.

On 2026-09-21, the supplied export validated against all 63 records. Its 68 events ended with 10 verified, 46 proposed and 7 needs-review UI decisions. The stricter multi-reviewer audit found 9 single-review and 54 pending records, zero agreed. After explicit user confirmation, 47 nonempty notes were promoted into the new draft. These are not 47 independently approved ground-truth corrections. The original history and all geometry concerns remain available for review.

Diagnostic run without repeat review

Update: the user explicitly authorized publishing the complete input ZIP. The public training/kaggle_body_dev_diagnostic.ipynb now automatically downloads the archive from a pinned GitHub commit and verifies SHA-256. Import only this small notebook into Kaggle, enable Internet and GPU, then Run All. No manual dataset upload is needed. The private embedded-notebook method below remains an alternative, but its file is too large for Kaggle's 1 MB notebook-import limit.

The user explicitly chose to skip another review. Draft status is retained; this does not create approved benchmark labels. Build a PRIVATE Kaggle notebook:

python -m training.build_private_body_notebook --manifest data/body-dev-corrected-draft-v1/manifest.jsonl --output data/body-dev-kaggle-v2

Import kaggle_body_dev_PRIVATE.ipynb from the output directory into Kaggle, keep it private, enable GPU and Internet, and Run All. Images and corrected draft text are embedded; no separate dataset upload is needed. Do NOT commit this generated notebook to GitHub. The builder and runner contain no images.

Both pinned TrOCR models use the same 63 inputs. Report all-draft metrics and the 56-line subset excluding the seven needs-review decisions separately. The latter is still provisional, not an independently verified subset. Failed predictions stay in denominators. Report case-sensitive NFC/whitespace CER/WER and an additional lowercase CER, without historical spelling modernization. The result ZIP contains predictions, draft references, environment and hashes, not weights or reviewer names. No inference quality is claimed until Kaggle runs.