| --- |
| pretty_name: PaddleOCR-VL Whole-page Table Evaluation (2864 pages) |
| language: |
| - en |
| - zh |
| tags: |
| - document-understanding |
| - table-recognition |
| - paddleocr-vl |
| - teds |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: "shard_*.parquet" |
| --- |
| |
| # PaddleOCR-VL whole-page table evaluation |
|
|
| **2,864 full-page images, 4,037 ground-truth tables, 16 pixel-balanced Parquet shards.** |
| Only pages containing at least one published GT table are included. This is evaluation-only data, separate from the cropped-table benchmark [table_recognition_eval_canonical](https://huggingface.co/datasets/CheneyWong/table_recognition_eval_canonical). |
|
|
| | Benchmark | Pages | GT tables | |
| |---|---:|---:| |
| | CC-OCR: all four table recognition tasks | 300 | 300 | |
| | OmniDocBench 1.5 | 351 | 512 | |
| | OmniDocBench 1.6 | 458 | 665 | |
| | Real5: Scanning | 351 | 512 | |
| | Real5: Warping | 351 | 512 | |
| | Real5: Screen Photography | 351 | 512 | |
| | Real5: Illumination | 351 | 512 | |
| | Real5: Skew | 351 | 512 | |
|
|
| ## Format |
|
|
| `image.bytes` embeds the unchanged original encoded page; `image.path` is its filename. `benchmark` + `sample_id` is the unique key. `width`, `height`, `pixels` and `image_sha256` describe and verify each image. |
|
|
| `gt_tables_json` contains only published table annotations, including HTML, polygons and annotation IDs where available. `gt_table_html` is the list of table HTML strings. `gt_page_json` preserves the original page annotation for provenance. CC-OCR has no published table-position annotation in this bundle: its polygon is null, and its original HTML is retained in `gt_html`. |
|
|
| GT must never enter the inference request. Parse the whole image with PaddleOCRVL, extract table blocks from the prediction, then match and score with the appropriate official OmniDocBench/CC-OCR implementation. Report each benchmark separately; TEDS-S is primary and TEDS auxiliary. Missing GT tables must remain in the denominator. |
|
|
| ```python |
| from datasets import load_dataset, Image |
| pages = load_dataset("CheneyWong/table_recognition_eval_fullpage", split="test") |
| pages = pages.cast_column("image", Image()) |
| ``` |
|
|
| ## Balance and provenance |
|
|
| Shards use longest-processing-time allocation by original `width * height`, without resizing or re-encoding. Their maximum/minimum total pixel ratio is 1.0000567611. `pack_audit.json` records counts and SHA256 for every shard; `index.jsonl` records each row's shard and position. |
|
|
| Source datasets retain their original provenance and terms. This bundle does not alter their annotations or claim ownership of their content. It contains no model predictions and is not a training split. |
|
|
|
|
| ## Per-page multiple-table predictions |
|
|
| Each Parquet row remains **one full page**, even when it contains multiple GT tables or predictions. The lists do not imply one-to-one matches or index alignment. |
|
|
| - `gt_table_ids`, `gt_table_html`, `gt_table_otsl`: every GT table on the page. IDs follow the original GT annotation-list order; original annotation IDs, order and polygons remain in `gt_tables_json`. |
| - `pred_table_ids`, `pred_table_html`, `pred_table_otsl`: every extracted predicted table on the page, retaining duplicates as distinct table instances. |
| - `pred_tables_json`: per-table parsing block ID, within-block index, bbox, polygon, associated layout detection indices/scores, unchanged original block/HTML and derived OTSL. |
| - `layout_table_regions_json`: all PP-DocLayoutV3 table detections and table parsing blocks, including detections without an associated complete HTML result. |
| - `prediction_model`: **PaddleOCR-VL-1.6 base**. These cached baseline predictions are not outputs of our SFT/RL checkpoints. |
| - `otsl_provenance`: GT and prediction OTSL are independently derived from their respective HTML, not recovered raw VLM tokens. |
|
|
| This update includes 4,502 baseline predicted tables across 2,864 pages, with 658 pages containing multiple predicted tables. The 4,037 GT tables are stored separately from predictions. Official evaluation still requires matching tables; never pair GT and predictions by array index. Images, page membership, pixel-based shard assignment and official metric results are unchanged. No inference or scoring was rerun to create these added columns. |
|
|
| ## PaddleOCR-VL-1.6 baseline evaluation JSON |
|
|
| [paddleocrvl1.6_eval.json](./paddleocrvl1.6_eval.json) is the single aggregate baseline JSON for this track. Completed: 2,864 GT-table pages and 4,037 GT tables. Official per-source summaries and all official per-page table matches are included. |
|
|
| See [EVALUATION.md](./EVALUATION.md) for both entry scripts, commands, scope, completion status and environments. Parquet image/GT bytes are unchanged by this update. |
|
|
