ko-hand-ocr 0.2.0__tar.gz

ko_hand_ocr-0.2.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

ko_hand_ocr-0.2.0/NOTICE

@@ -0,0 +1,20 @@
+ko-hand-ocr
+Copyright 2026 ko-hand-ocr contributors
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+이 저작물에는 아래 것이 들어 있습니다.
+
+- 인코더 가중치는 `facebook/deit-small-patch16-224` (Apache-2.0, Meta Platforms, Inc.)
+  에서 시작해 미세조정한 것입니다.
+- 디코더 가중치는 이 저장소에서 처음부터 학습했습니다.
+- 학습 이미지는 아래 글꼴로 그려 만들었습니다.
+    · 나눔손글씨 109벌 (NAVER Corporation / CLOVA AI OCR Team)
+    · 한글 손글씨 글꼴 14벌 (google/fonts, SIL Open Font License 1.1)
+  **글꼴 파일은 이 저장소에 들어 있지 않으며 재배포하지 않습니다.**
+  가중치는 글자 모양의 저작물이 아니라 그림에서 배운 값입니다.
+  자세한 내용은 PROVENANCE.md 를 보십시오.

ko_hand_ocr-0.2.0/PKG-INFO

@@ -0,0 +1,322 @@
+Metadata-Version: 2.4
+Name: ko-hand-ocr
+Version: 0.2.0
+Summary: Korean handwriting OCR - reads one line of handwritten Korean/English. 31M params, CPU-only, no dataset terms inherited.
+License-Expression: Apache-2.0
+Project-URL: Source, https://github.com/jysvai/ko-hand-ocr
+Project-URL: Issues, https://github.com/jysvai/ko-hand-ocr/issues
+Keywords: ocr,handwriting,korean,hangul,trocr
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: torch>=2.5
+Requires-Dist: transformers>=4.45
+Requires-Dist: Pillow>=10.0
+Provides-Extra: train
+Requires-Dist: numpy>=1.26; extra == "train"
+Dynamic: license-file
+
+<div align="center">
+
+# ko-hand-ocr
+
+**Reads one line of handwritten Korean + English.**
+A 31M model trained from scratch on synthetic data — no handwriting dataset, no inherited terms.
+
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/jysvai/ko-hand-ocr/blob/main/LICENSE)
+[![Python](https://img.shields.io/badge/python-3.10%2B-blue)](https://github.com/jysvai/ko-hand-ocr)
+[![Params](https://img.shields.io/badge/params-31M-1baf7a)](https://github.com/jysvai/ko-hand-ocr)
+[![Weights](https://img.shields.io/badge/weights-116MB-1baf7a)](https://github.com/jysvai/ko-hand-ocr/releases/tag/v0.2.0)
+[![CPU](https://img.shields.io/badge/runs%20on-CPU%20only-eda100)](https://github.com/jysvai/ko-hand-ocr)
+[![Data](https://img.shields.io/badge/training%20data-synthetic-e87ba4)](https://github.com/jysvai/ko-hand-ocr/blob/main/PROVENANCE.md)
+
+[한국어](https://github.com/jysvai/ko-hand-ocr/blob/main/README.md) · **English**
+
+</div>
+
+```
+photo  ->  line cutting  ->  ko-hand-ocr  ->  "부서 : 포테토뭉부서"
+```
+
+---
+
+## At a glance
+
+| | |
+|---|---|
+| **What it reads** | One line of handwritten Korean + English, from a photo |
+| **Model** | ViT-Small encoder (ImageNet, Apache-2.0) + 4-layer TrOCR decoder trained from scratch |
+| **Vocabulary** | 229 jamo tokens — not 11,172 syllables, so unseen characters are still writable |
+| **Parameters** | **31M** — one seventh of `ko-trocr` (213.7M) |
+| **Weights** | **116 MB** single · 465 MB four-way ensemble *(downloaded separately)* |
+| **Package** | 83 KB — the code only |
+| **Speed** | **0.16 s per cell** · 1.37 s for a whole photo — **CPU only, no GPU** |
+| **Memory** | 890 MB single · 1.35 GB ensemble |
+| **Accuracy** | **93.5%** on handwriting fonts never seen in training (93.9% ensemble) |
+| **Training data** | Synthetic — drawn on the fly from OFL fonts. Nothing is stored on disk |
+| **License** | Apache-2.0, weights included — **no dataset terms inherited** |
+| **Python** | 3.10+ · PyTorch 2.5+ |
+
+## What it does — and does not
+
+**Does**
+
+- Reads a **whole photo**: finds the lines, cuts them, reads each one (`read_photo`)
+- Reads **pre-cut line images**, batched (`read`)
+- **Constrained decoding** — restrict the output to a candidate list, and report whether the free and constrained readings agree (`both`)
+- **Ensemble reading** — run several checkpoints and take the answer they agree on
+- Runs **fully offline on CPU.** Nothing is sent anywhere
+
+**Does not**
+
+- Split a table cell into label and value — `read_photo` only cuts down to lines
+- Handle vertical writing or merged cells
+- Stay silent on an empty cell — a cell that is pure scribble still produces something
+
+## Why it exists
+
+In practice the only publicly available Korean handwriting OCR model is
+`ddobokki/ko-trocr`, and its training data comes from AI Hub, which places conditions
+on purpose of use and on redistribution. That blocks it from being embedded in an
+in-house tool or shipped as a public package. So **the same capability was rebuilt
+from scratch** — with a provenance chain that can be audited part by part.
+
+![ko-hand-ocr vs ko-trocr](https://raw.githubusercontent.com/jysvai/ko-hand-ocr/main/bench-compare.svg)
+
+## Install
+
+```bash
+pip install git+https://github.com/jysvai/ko-hand-ocr.git
+```
+
+> Not on PyPI yet. Until it is, install from the repository.
+
+Weights ship separately — they are too large for the repository, so they are attached
+to [Releases](https://github.com/jysvai/ko-hand-ocr/releases/tag/v0.2.0).
+
+| Download | Size | What it is |
+|---|---|---|
+| `ko-hand-ocr-v46.zip` | 116MB | **A single checkpoint.** Enough for most uses. 0.16 s/cell |
+| `ko-hand-ocr-ensemble.zip` | 465MB | Four checkpoints. More accurate, 10x slower |
+
+```bash
+curl -LO https://github.com/jysvai/ko-hand-ocr/releases/download/v0.2.0/ko-hand-ocr-v46.zip
+```
+
+Pass the unzipped folder straight to `Reader()`. It must contain `config.json`,
+`vocab.json` and `model.safetensors`.
+
+## Quick start
+
+```python
+from kohandocr.reader import Reader
+
+reader = Reader("ko-hand-ocr-v46", device="cpu")   # the unzipped folder
+
+# A whole photo. Line cutting is included.
+reader.read_photo(open("scan.jpg", "rb").read())
+# -> ['보안점검표', '부서 : 포테토뭉부서', '이름 : 감자밭']
+
+# If the lines are already cut, hand the images in directly. Batch them for speed.
+reader.read([cell_a, cell_b, cell_c])
+
+# Given a candidate list, the output can be constrained to it.
+reader.both([cell_b], options=["포테토뭉부서", "감자밭", "김클로드"])
+# -> [{'text': '부서 : 포테토뭄부서', 'constrained': '포테토뭉부서', 'agrees': True}]
+```
+
+**Do not trust the constrained result on its own.** Because it is constrained, the
+model *must* return something from the list. If the true answer is not in the list you
+get a confident wrong answer and nothing looks off. When `agrees` is false, hand it
+to a human.
+
+### Reading with several checkpoints
+
+Different checkpoints fail on different cells. Train a little longer on the same data
+and some cells survive while others die — which tells you those cells were being read
+by a thin margin. So keep several checkpoints and take **the answer they agree on**.
+
+```
+unzipped/
+  config.json  vocab.json  model.safetensors
+  also/
+    v20b/  config.json  vocab.json  model.safetensors
+    v19/   ...
+```
+
+If `also/` is present, `Reader` uses it automatically. The call site does not change.
+
+Combining by confidence was measured and **it did not work** — the model is sometimes
+more confident about a wrong answer. So selection is by **how much the outputs agree**,
+not by confidence.
+
+## How it is built
+
+```
+  64 x 640 line image
+          |
+   ViT-Small encoder          facebook/deit-small-patch16-224 (ImageNet-1k, Apache-2.0)
+   patch 16                   pretrained weights, continued at a lower learning rate
+          |
+   TrOCR decoder              4 layers, 6 heads — trained from scratch
+          |
+   229 jamo tokens            ㄱ ㅏ ㅁ ... assembled into 감
+          |
+      "부서 : 감자밭"
+```
+
+Two decisions carry most of the result.
+
+**Jamo, not syllables.** Memorising all 11,172 Hangul syllables needs a large output
+layer and still fails on anything rare. 229 jamo compose into any syllable, so a
+character the model never saw in training is still writable — which matters for names.
+
+**The encoder is not frozen, but it is not shaken either.** It already learned to see
+on ImageNet, so it continues at a lower learning rate while the blank decoder learns
+fast. Training both at the same rate destroys what the encoder knew.
+
+## Accuracy
+
+Measured two ways. **Both matter.** All figures below come from
+`python tools/verify.py` (400 lines per font, 11 photos).
+
+| | unseen fonts (mean) | worst font | 11 handwriting photos | worst photo |
+|---|---|---|---|---|
+| single (v46) | 93.5% | 86.4% | 89.6% | 66.7% |
+| four | **93.9%** | **88.1%** | **95.6%** | **83.1%** |
+
+Similarity is measured per jamo; on the photo side each photo gets one vote.
+
+**Do not take the photo number at face value.** It is measured on 11 photos (37 cells)
+taken by the author, so it swings ±5%p. The figure closer to what you would see on
+someone else's handwriting is the **unseen-fonts** column — tested only on six font
+families never used in training, where the swing is half as wide (±2-3%p).
+
+Four checkpoints open a **large gap on photos** (95.6 vs 89.6) and a small one on
+fonts (93.9 vs 93.5). The photo gap is large because where one checkpoint collapses
+another carries the cell — the worst photo goes from 66.7% with one checkpoint to
+83.1% with four. The cost is being ten times slower.
+
+### Per font (400 lines each)
+
+![Accuracy by handwriting font](https://raw.githubusercontent.com/jysvai/ko-hand-ocr/main/bench-accuracy.svg)
+
+Look only at the average and **the font that collapses is hidden.** The real spread
+is 12%p.
+
+| Font | single (v46) | four |
+|---|---|---|
+| **EastSeaDokdo (동해독도)** | **86.4%** | **88.1%** |
+| KirangHaerang (기랑해랑) | 92.4% | 92.6% |
+| HiMelody (하이멜로디) | 93.2% | 93.2% |
+| NanumPenScript (나눔펜스크립트) | 94.0% | 94.2% |
+| GamjaFlower (감자꽃) | 96.9% | 97.0% |
+| SingleDay (싱글데이) | 97.9% | 98.1% |
+
+What has to clear the bar is not the mean but **the lowest font.** The goal is 90% on
+all six; EastSeaDokdo has not got there yet.
+
+All six are fonts **never used in training**, and all are SIL Open Font License.
+
+```bash
+python tools/holdout.py <checkpoint-dir> --per-font --sample sample.png
+```
+
+### What the test sheet actually looks like
+
+![unseen-font test sheet](https://raw.githubusercontent.com/jysvai/ko-hand-ocr/main/holdout-sample.png)
+
+Three lines per font, with the ground truth, what the model read, and the similarity.
+The pale band on the right is the leftover space from fitting into a 64x640 frame —
+this is exactly what the model sees.
+
+It shows why the gap opens. **SingleDay** has well-separated characters and intact
+strokes, close to print. **EastSeaDokdo** smears and drops strokes, so `안전사무국`
+looks like a different word almost entirely — a human cannot read it without context
+either.
+
+This is **an image, not a font file.** Drawing glyphs with a font is what the OFL
+permits; what is not redistributed is the `.ttf` files themselves
+([PROVENANCE.md](https://github.com/jysvai/ko-hand-ocr/blob/main/PROVENANCE.md)).
+
+## Speed and footprint (measured)
+
+No GPU needed. Emitting jamo one at a time is the bottleneck, so more cores do not help
+much — and by the same token **it does not get slower on a weak machine.**
+
+CPU only, median over 11 photos (37 cells). Time is split into **cutting and reading** —
+shrinking the model does not shrink cutting, and when a photo holds only three or four
+lines, cutting is more than half the cost.
+
+| | cut | read | one photo (3.4 cells) | per cell |
+|---|---|---|---|---|
+| single, beam 5 | 0.82s | 0.55s | 1.37s | 0.16s |
+| single, greedy | 0.82s | 0.32s | 1.13s | 0.10s |
+| four, beam 5 | 0.82s | 5.73s | 6.55s | 1.70s |
+
+Memory is about 890MB for one checkpoint, about 1.35GB for four.
+
+Produced with `python tools/bench.py --all --device cpu`. That harness walks exactly the
+path the application walks — a speed measured along a different path is not the speed
+the user gets.
+
+## Where it runs
+
+**Built and tested on Windows.** Training and inference were both run on
+Windows 11 with Python 3.13/3.14.
+
+| | |
+|---|---|
+| Windows | Where it was built. Training and inference both verified |
+| macOS | **Untested.** Inference is pure PyTorch + PIL so it should run, but it has not been confirmed |
+| Linux | Not tried yet |
+
+`tools/train.ps1` is PowerShell, so it is Windows-only. Elsewhere, call
+`python -m kohandocr.train` directly — all that script does is wait for the previous
+run to release the GPU, launch it, and watch the first few steps.
+
+## Training
+
+No data is baked ahead of time. A corpus line is generated, drawn with a handwriting
+font, and fed in **on the spot**. Nothing is left on disk, and a change to the
+distribution takes effect from the next step.
+
+```bash
+python tools/fetch_fonts.py                       # fetch OFL handwriting fonts
+python -m kohandocr.train --out runs/v1 \
+       --steps 20000 --batch 40 --workers 3 --letters 128
+```
+
+Fonts are looked up in this order. **No path is hard-coded.**
+
+1. whatever `--fonts` was given
+2. the `KOHAND_FONTS` environment variable
+3. `~/.cache/ko-hand-ocr/fonts`
+
+The fonts are not in this repository and are **not redistributed.** The 109 Nanum
+handwriting fonts come from [clova.ai/handwriting](https://clova.ai/handwriting).
+Details in [PROVENANCE.md](https://github.com/jysvai/ko-hand-ocr/blob/main/PROVENANCE.md).
+
+How closely the synthetic images match real handwriting is checked against the measured
+distributions recorded in `kohandocr/measured.json` — nine of them (ink coverage,
+density, aspect ratio, margins and so on).
+
+```bash
+python tools/match.py        # target vs current synthesis, side by side
+```
+
+**Running longer does not make it better.** After changing the synthesis distribution,
+run short, keep intermediate checkpoints, and pick between them (measured: 15,000 steps
+was the peak and 30,000 steps was worse).
+
+```bash
+python tools/pick.py runs/v1-15000 runs/v1-30000    # score the kept checkpoints and pick
+```
+
+## License
+
+Apache-2.0, weights included. The licenses of the fonts used in training do not attach
+to the weights — the weights are not a derivative of the glyph artwork, they are values
+learned from images. The reasoning is recorded part by part in
+[PROVENANCE.md](https://github.com/jysvai/ko-hand-ocr/blob/main/PROVENANCE.md).