Datasets:
Formats:
parquet
Size:
100K - 1M
Tags:
table-recognition
table-structure-recognition
document-image-analysis
ocr
document-understanding
cv
License:
Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -248,11 +248,6 @@ machine with many cores.
|
|
| 248 |
|
| 249 |
- **TABLET train pool** (17,720 rows) — replaced by the 909-row test split,
|
| 250 |
since the train pool overlaps the evaluation set.
|
| 251 |
-
- **OCRBench-v2 rows without a bbox** (542) — their prompt says "all tables
|
| 252 |
-
in the image", so there is no region to crop to. Only the 158 rows with an
|
| 253 |
-
explicit bbox are included.
|
| 254 |
-
- **OmniDocBench v1.6** (665) — no public annotation file ships the table
|
| 255 |
-
HTML, so the images alone would be unscorable.
|
| 256 |
- **OCRBench-v2 non-table tasks** (9,300) — OCRBench-v2 is a general document-VQA
|
| 257 |
benchmark. POIE, ChartQA, MMMU, textvqa and the rest carry no table annotation
|
| 258 |
at all, so they cannot score a table recogniser. Of the 700 genuinely
|
|
|
|
| 248 |
|
| 249 |
- **TABLET train pool** (17,720 rows) — replaced by the 909-row test split,
|
| 250 |
since the train pool overlaps the evaluation set.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 251 |
- **OCRBench-v2 non-table tasks** (9,300) — OCRBench-v2 is a general document-VQA
|
| 252 |
benchmark. POIE, ChartQA, MMMU, textvqa and the rest carry no table annotation
|
| 253 |
at all, so they cannot score a table recogniser. Of the 700 genuinely
|