| --- |
| license: apache-2.0 |
| pretty_name: Table Recognition Benchmark (176,810 crops; native-page benchmarks separated) |
| tags: [table-recognition, table-structure-recognition, ocr, teds, paddleocr-vl] |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: "shard_*.parquet" |
| --- |
| |
| # Canonical table-crop evaluation |
|
|
| **176,810 table crops in 64 pixel-balanced shards.** Evaluation only. |
|
|
| OmniDocBench 1.5, OmniDocBench 1.6 and CC-OCR have been physically removed from these shards. Their full-page evaluation, together with Real5, lives in the separate [whole-page dataset](https://huggingface.co/datasets/CheneyWong/table_recognition_eval_fullpage). Real5 was never part of this crop pack. TabRecSet zh/en and audited invalid annotations are excluded. |
|
|
| Original encoded images are unchanged. Shards are rebuilt using LPT by `width * height`, not by row count. `release_audit.json` records per-source exclusions, GT patches, shard counts, pixel loads and SHA256. |
|
|
| Existing schema is preserved (`bytes`, `source`, `benchmark`, `sample_id`, `rows`, `cols`, `otsl`, `html`, `resolution`, `width`, `height`, `structure_only`, `otsl_prev`, `otsl_changed`). Additional columns: `canonical_sample_id` retains the pre-repack `shard#rowgroup#row` identity; `image_sha256` verifies original bytes; `gt_repair_provenance` identifies an audited GT override when applied. |
|
|
| Do not join legacy results using the new physical row positions or empty upstream `sample_id` values. Use `canonical_sample_id` to recover the original pack identity. No model inference or scoring was rerun for this release. |
|
|
| TEDS-S is primary. For GT without cell text, report TEDS as N/A and score structure only. The original historical README and Git history retain the previous 184,074-row release description. Auxiliary legacy files outside `shard_*.parquet` are not part of the default test split. |
|
|
| | Source | Crops | |
| |---|---:| |
| | fintabnet | 10,397 | |
| | hitab | 3,461 | |
| | mmsci | 12,129 | |
| | ocrbenchv2 | 158 | |
| | pubtables1m | 93,834 | |
| | pubtablesv2 | 7,680 | |
| | pubtabnet_icdar21 | 9,064 | |
| | scitsr | 3,000 | |
| | sprint_mustard | 1,429 | |
| | synthtabnet | 29,799 | |
| | tablet | 909 | |
| | tableverse | 4,950 | |
|
|
| ## PaddleOCR-VL-1.6 baseline evaluation JSON |
|
|
| [paddleocrvl1.6_eval.json](./paddleocrvl1.6_eval.json) is the single aggregate baseline JSON for this track. Historical base inference covered the former 184,074-row release. Current 176,810-row scores remain null until the repaired-GT evaluation completes; historical metrics are explicitly separated under `historical_evaluation` and have no verified current-GT binding. |
|
|
| See [EVALUATION.md](./EVALUATION.md) for both entry scripts, commands, scope, completion status and environments. Parquet image/GT bytes are unchanged by this update. |
|
|
