Results: add NAVER CLOVA OCR (General + table extraction), v3 total 0.853
Browse files
README.md
CHANGED
|
@@ -49,7 +49,7 @@ table slices use `teds`, text slices `1 - cer`). v3 raises tables and handwritin
|
|
| 49 |
A slice an engine has no rows for is dropped and the remaining weights are renormalised (row marked `partial`).
|
| 50 |
The v2 weights (tables .44 · text .36 · handwriting .10 · order .10) stay available as `scorer/koocr_total_v2.py`.
|
| 51 |
|
| 52 |
-
## Results (v3 total, 2026-11-01)
|
| 53 |
|
| 54 |
Scored on all 1,070 pages (public + holdout). Higher is better. `tbl_form` / `tbl_text` split the grid-table
|
| 55 |
slices into grid shape and cell text; they are shown, not added to the total.
|
|
@@ -61,6 +61,7 @@ slices into grid shape and cell text; they are shown, not added to the total.
|
|
| 61 |
| schift-ocr-1-a1b — private | 0.868 | 0.864 | 0.873 | 0.843 | 0.928 | 0.870 | 0.831 |
|
| 62 |
| schift-ocr-1 — private | 0.862 | 0.862 | 0.869 | 0.823 | 0.928 | 0.873 | 0.842 |
|
| 63 |
| datalab-to/chandra-ocr-2 | 0.860 | 0.886 | 0.804 | 0.816 | 0.933 | 0.870 | 0.850 |
|
|
|
|
| 64 |
| Qwen/Qwen3.8-27B-FP8 | 0.852 | 0.883 | 0.862 | 0.724 | 0.931 | 0.868 | 0.850 |
|
| 65 |
| PaddlePaddle/PaddleOCR-VL-1.6 | 0.851 | 0.879 | 0.814 | 0.755 | 0.976 | 0.872 | 0.864 |
|
| 66 |
| MinerU 2.5 (local pipeline) | 0.758 | 0.804 | 0.810 | 0.478 | 0.984 | 0.759 | 0.750 |
|
|
@@ -70,6 +71,12 @@ slices into grid shape and cell text; they are shown, not added to the total.
|
|
| 70 |
|
| 71 |
`private` = our models, not released.
|
| 72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
## Holdout
|
| 74 |
|
| 75 |
20% of every slice is held out and is not in this repo: a page is held out when `int(sha1(id), 16) % 5 == 0`
|
|
|
|
| 49 |
A slice an engine has no rows for is dropped and the remaining weights are renormalised (row marked `partial`).
|
| 50 |
The v2 weights (tables .44 · text .36 · handwriting .10 · order .10) stay available as `scorer/koocr_total_v2.py`.
|
| 51 |
|
| 52 |
+
## Results (v3 total, 2026-11-01; CLOVA added 2026-11-09)
|
| 53 |
|
| 54 |
Scored on all 1,070 pages (public + holdout). Higher is better. `tbl_form` / `tbl_text` split the grid-table
|
| 55 |
slices into grid shape and cell text; they are shown, not added to the total.
|
|
|
|
| 61 |
| schift-ocr-1-a1b — private | 0.868 | 0.864 | 0.873 | 0.843 | 0.928 | 0.870 | 0.831 |
|
| 62 |
| schift-ocr-1 — private | 0.862 | 0.862 | 0.869 | 0.823 | 0.928 | 0.873 | 0.842 |
|
| 63 |
| datalab-to/chandra-ocr-2 | 0.860 | 0.886 | 0.804 | 0.816 | 0.933 | 0.870 | 0.850 |
|
| 64 |
+
| NAVER CLOVA OCR (General + table extraction, API) | 0.853 | 0.874 | 0.736 | 0.943 | 0.800 | 0.884 | 0.875 |
|
| 65 |
| Qwen/Qwen3.8-27B-FP8 | 0.852 | 0.883 | 0.862 | 0.724 | 0.931 | 0.868 | 0.850 |
|
| 66 |
| PaddlePaddle/PaddleOCR-VL-1.6 | 0.851 | 0.879 | 0.814 | 0.755 | 0.976 | 0.872 | 0.864 |
|
| 67 |
| MinerU 2.5 (local pipeline) | 0.758 | 0.804 | 0.810 | 0.478 | 0.984 | 0.759 | 0.750 |
|
|
|
|
| 71 |
|
| 72 |
`private` = our models, not released.
|
| 73 |
|
| 74 |
+
NAVER CLOVA OCR: General OCR V2 API with `enableTableDetection: true`, `lang: ko`, run 2026-11-09, score only (commercial API, predictions not published).
|
| 75 |
+
Output is the API's words in API order (`lineBreak` ends a line); each detected table is written once as HTML `<table>`
|
| 76 |
+
(colspan/rowspan from the API) where its first word falls, and words inside a table box are not repeated.
|
| 77 |
+
Three handwriting photos over the API's 8,000 px limit were shrunk to 8,000 px on the long side.
|
| 78 |
+
Its low `text` comes mostly from suneung2026: the API returns two-column exam pages as one large table, so the two columns interleave.
|
| 79 |
+
|
| 80 |
## Holdout
|
| 81 |
|
| 82 |
20% of every slice is held out and is not in this repo: a page is held out when `int(sha1(id), 16) % 5 == 0`
|