Datasets:

CheneyWong
/

table_recognition_datasets_mixed

+---
+license: other
+task_categories:
+  - image-to-text
+language:
+  - eng
+  - zho
+tags:
+  - table-recognition
+  - ocr
+  - otsl
+  - parquet
+  - evaluation
+---
+# PaddleOCR-VL Table Recognition Evaluation Pack
+Table-recognition evaluation data for PaddleOCR-VL, packaged as **32 self-contained
+parquet shards** with images embedded inline.
+Every table is stored twice — as **OTSL** and as **canonical HTML** — so it can be
+scored against either dialect without re-deriving one.
+## Folders in this repo
+| Folder | Rows | Images | Use it for |
+|---|---|---|---|
+| `table_eval_paddleocr_vl/` | 240,075 | 152,982 | **Default.** Blocklist already applied |
+| `table_eval_paddleocr_vl_excluded_toolong/` | 3,830 | ~3,635 | The samples withheld from evaluation because their tables are too long |
+## Why this exists
+The source of truth was 49 *cohort* JSONL files pointing at loose image files. Three
+problems made that untenable:
+1. **Inodes.** `/autodl-fs/data` is capped at 200,000 inodes and ran at 97%. One file
+   per image is the root cause; sharding collapses 240,075 rows into 32 files.
+2. **Fragility.** 240,417 sample references pointed at `/tmp/povl_v5_*` and
+   `/dev/shm/*` — both volatile, wiped on every container restart. Images are now
+   embedded and snapshotted.
+3. **Duplicates.** 51,251 rows were the same sample appearing in more than one cohort.
+## Image recovery
+164,055 images were recovered by streaming the upstream archives — **100 % hit rate**,
+no partial coverage:
+| Source | Recovered | Time |
+|---|---|---|
+| PubTables-1M (`PubTables-1M_image_html_test.tar.gz`) | 93,656 / 93,656 | 15 s |
+| SynthTabNet (`SynthTabNet_png_html.tar`) | 43,433 / 43,433 | 388 s |
+| FinTabNet-OTSL (`test.zip`) | 10,197 / 10,197 | 2 s |
+| PubTables-v2 (`PubTables-v2_image_html_test.tar.gz`) | 7,663 / 7,663 | 5 s |
+`data/packs/` was *not* usable for this — it holds a 30,000-sample random subset of
+PubTables-1M **train**, which has zero overlap with the test split the cohorts
+reference.
+87,093 rows (36 %) still carry `has_image = false`. They are the degraded/derived sets
+(`ooddeg`, `opdpool`, `val10k`, `deg`, `CXRender2`) that lived only in `/dev/shm`
+with no surviving upstream archive. Their labels are intact and will work again once
+the images are regenerated.
+## Blocklist — why this folder is already filtered
+Samples whose tables are too long, or whose ground truth is unusable, are excluded at
+**build time** rather than in the scoring script, so the exclusion is auditable:
+| Rule | Count | Source |
+|---|---|---|
+| `table_too_long` | 2,271 dropped | `*_test_union_excluded_ids.jsonl` |
+| `unusable_gt` | 2 | `sample_blocklist.json` |
+Dropping is **not** silent. The withheld rows live in
+`table_eval_paddleocr_vl_excluded_toolong/`, flagged with an `excluded_reason` column.
+Note the id formats differ between the lists and the cohorts
+(`FinTabNet_OTSL_test::X` vs `fintabnet::X`). `povl_v5/src/eval_blocklist.py` performs
+the mapping; comparing the raw strings yields a false "no overlap".
+## Schema
+| column | type | description |
+|---|---|---|
+| `sample_id` | string | Stable id, byte-identical to the source cohort JSONL |
+| `cohort` | string | Primary cohort |
+| `cohorts` | list\<string\> | **All** cohorts this row belongs to (post-dedupe) |
+| `n_cohorts_merged` | int32 | `len(cohorts)`; > 1 means duplicates were merged |
+| `dataset` | string | e.g. `PubTables-1M`, `SynthTabNet` |
+| `origin` | string | `selfbuilt` / `ood` / `public` |
+| `group` | string | `pub` / `ood` / `cx` |
+| `image` | binary | Original encoded bytes, **never re-encoded** |
+| `img_format` | string | `jpeg` / `png` |
+| `has_image` | bool | `false` when the upstream image is gone |
+| `img_w`, `img_h` | int32 | Columns, so eval never decodes to balance load |
+| `megapixels` | float32 | `w*h/1e6` |
+| `res_bucket` | string | `<0.5MP` / `0.5-1MP` / `1-2MP` / `2-4MP` / `>=4MP` |
+| `gt_otsl` | string | OTSL in the model's six-token dialect |
+| `gt_html` | string | Canonical HTML |
+| `gt_format_native` | string | Which of the two is authoritative |
+| `metrics` | list\<string\> | e.g. `["teds_s"]` when only structure matters |
+| `n_rows`, `n_cols`, `n_cells` | int32 | R / C / N |
+| `n_rowspan`, `n_colspan` | int32 | Cells with span > 1 |
+| `max_rowspan`, `max_colspan` | int32 | Largest span in the table |
+| `n_merge_cells` | int32 | M |
+| `complexity` | string | `simple` / `medium` / `complex` |
+| `tags` | list\<string\> | `long`, `wide`, `merge_dense`, … |
+| `numeric_cell_frac` | float32 | Fraction of non-empty cells containing a digit |
+| `has_numeric` | bool | `numeric_cell_frac >= 0.10` |
+| `img_sha1` | string | Content hash of the image bytes |
+| `gt_conflict` | bool | Same image carried a conflicting label |
+| `otsl_legal` | bool | Passed `check_otsl_strict` |
+| `otsl_roundtrip_ok` | bool | OTSL re-parses to the identical grid |
+**Read `gt_format_native`, not `gt_otsl` blindly.** OTSL→HTML is lossy for tables with
+uncovered holes: `cells_to_otsl` writes `<ecel>` at each hole and re-parsing counts
+every one as its own cell, inflating N. `otsl_roundtrip_ok` flags the affected rows.
+## Deduplication
+**Every row appears exactly once.** Rows are keyed by `sample_id` *and* by
+`(img_sha1, gt_otsl)`; on collision the duplicate is dropped and its cohort appended to
+the survivor's `cohorts` list. 291,917 references collapsed to 240,075 rows via 51,251
+exact merges, without losing cohort membership information.
+591 rows hit the **conflict** path — the same image with two different labels. They are
+flagged `gt_conflict=true` and recorded in `DEDUP_CONFLICTS.jsonl`. They cluster in
+`ood_oodtrain`, where a multi-table newspaper page was never cropped to the individual
+table, so one page image was paired with the GT of two unrelated tables. That is an
+annotation defect in the source cohort, not a packaging artifact.
+## Distributions
+### Table geometry
+| metric | min | p25 | p50 | p75 | p90 | p99 | max |
+|---|---|---|---|---|---|---|---|
+| Rows (R) | 1 | 7 | 12 | 20 | 32 | 49 | 132 |
+| Columns (C) | 1 | 4 | 5 | 8 | 12 | 15 | 115 |
+| Cells (N) | 1 | 30 | 62 | 119 | 182 | 345 | 5,060 |
+| Cells with rowspan>1 | 0 | 0 | 0 | 2 | 5 | 16 | 355 |
+| Cells with colspan>1 | 0 | 0 | 1 | 3 | 6 | 14 | 316 |
+| Max rowspan | 1 | 1 | 1 | 2 | 4 | 10 | 99 |
+| Max colspan | 1 | 1 | 2 | 4 | 6 | 11 | 51 |
+### Image geometry
+| metric | min | p25 | p50 | p75 | p90 | p99 | max |
+|---|---|---|---|---|---|---|---|
+| Width (px) | 103 | 416 | 512 | 608 | 653 | 895 | 957 |
+| Height (px) | 9 | 140 | 241 | 425 | 512 | 899 | 925 |
+| Megapixels | 0.001 | 0.066 | 0.123 | 0.205 | 0.277 | 0.708 | 0.810 |
+### Categorical
+| Origin | Count | Share |
+|---|---|---|
+| public | 190,034 (79.2%) | ood | 41,567 (17.3%) | selfbuilt | 8,474 (3.5%) |
+| `has_numeric` | Count | Share |
+|---|---|---|
+| True | 225,748 (94.0%) | False | 14,327 (6.0%) |
+| Resolution bucket | Count | Share |
+|---|---|---|
+| <0.5MP | 235,687 (98.2%) | 0.5-1MP | 4,388 (1.8%) |
+| Complexity | Count | Share |
+|---|---|---|
+| medium | 100,153 (41.7%) | complex | 93,554 (39.0%) | simple | 46,368 (19.3%) |
+| Native GT format | Count | Share |
+|---|---|---|
+| otsl | 218,601 (91.1%) | html | 21,474 (8.9%) |
+## Scoring
+Predictions and ground truth are normalised to canonical HTML before TEDS, so a model
+that answers in OTSL and one that answers in HTML are compared on equal terms.
+```python
+import sys; sys.path.insert(0, "povl_v5/src")
+import eval_mix
+rows = eval_mix.load_from_packs("data/eval_packs_filtered")
+```
+Or with the CLI:
+```bash
+python povl_v5/src/eval_mix.py --model DIR --tag run1 \
+    --packs povl_v5/data/eval_packs_filtered --gpus 0 1 2 3
+```
+Images resolve lazily per request through `eval_pack_reader.py`. Never call
+`pq.read_table(shard).to_pylist()` on the inference path — it re-decodes an entire row
+group per request and collapses under concurrency.
+For the SPRINT-MUSTARD family the GT is deliberately all-empty `<td>`, so **only TEDS-S
+is meaningful**; its TEDS column reads artificially low.