CheneyWong/table_recognition_eval_canonical

Table Recognition Benchmark — canonical PaddleOCR-VL 1.6 OTSL

17 public and self-built table sources merged into 64 uniformly sized shards (184,074 tables, ~8.6 GB of images). Every row carries an OTSL ground truth in PaddleOCR-VL 1.6's own OTSL convention, the HTML ground truth, and a flag marking the tables that carry no cell text at all.

This is a corrected release. The previous release (CheneyWong/table_recognition_eval) shipped an OTSL column whose rowspan structure did not follow PaddleOCR-VL's own parser, and whose documentation listed a token that does not exist in the 1.6 vocabulary. See What changed — in short, 42.6% of the rows have a different ground truth here.

Layout

shard_0000.parquet ... shard_0063.parquet   # 64 evenly-sized shards
scores_reference_base.parquet                # reference run: PaddleOCR-VL 1.6 (base), per-sample scores
scores_reference_sft_ckpt2800.parquet        # reference run: our SFT ckpt-2800, per-sample scores
README.md                                   # you are here
scripts/otsl_norm.py                        # the canonicaliser (HTML -> 1.6 OTSL)
scripts/canonical_teds.py                   # the scorer used for the reference runs
scripts/README.md                           # how to rebuild / verify

Shards are balanced by total image bytes, not row count — each holds ~169 MB of images, and every shard mixes 15–17 different sources, so a single shard already gives broad coverage.

Schema

column	type	meaning
bytes	binary	the table image, as stored (PNG/JPEG)
source	string	fine-grained original dataset name
benchmark	string	which of the 17 sources this row came from
sample_id	string	original sample identifier (may be empty)
rows / cols	int64	table grid dimensions
otsl	string	canonical OTSL ground truth (PaddleOCR-VL 1.6 convention)
html	string	HTML ground truth
resolution	string	WIDTHxHEIGHT
width / height	int64	image dimensions
structure_only	bool	true when the table has no cell text at all
otsl_prev	string	the ground truth of the previous release (kept for reference)
otsl_changed	bool	true when otsl differs from otsl_prev

What is OTSL

OTSL = Optimized Table Structure Language, the compact token language PaddleOCR uses to represent table structure. A table is a flat token stream — no nesting, no closing tags. The vocabulary is exactly six tokens (this is PaddleOCR-VL 1.6's own tokenizer, PaddleOCR-VL-1.6/tokenizer.json):

token	id	meaning
<fcel>	101309	a new (filled) cell starts here
<ecel>	101308	an empty cell
<lcel>	101311	continuation to the left (horizontal merge)
<ucel>	101312	continuation above (vertical merge)
<xcel>	101310	continuation both left and above
<nl>	101313	end of row

Cell text is inlined directly after its <fcel>, so the stream alternates structure and content:

<fcel>Year<fcel>Division<ecel><nl>
<fcel>2023<fcel>A<ecel><nl>

A rowspan is a continuation row, not a marker inside the cell:

<fcel>A<fcel>B<lcel><nl>     A spans two rows, B spans two columns
<ucel><fcel>C<fcel>D<nl>     <- the continuation row, <ucel> holds A's column

What changed, and why

The previous release's otsl column was produced by an HTML→OTSL converter that (a) put a rowspan's continuation marker inside the same row using a token <ulcel>, and (b) padded every row with <ecel> up to a fixed column count.

<ulcel> is not a PaddleOCR-VL 1.6 token. It is not in the model's vocabulary, so the model can never emit it; and the usual scorer (canonical_teds.py, used by OmniDocBench-style table evaluation) does not list it among the tokens it recognises either, so it silently ends up inside the cell text:

<fcel><ulcel><ulcel>characteristic   ->   cell text = "<ulcel><ulcel>characteristic"

Rows containing <ulcel> in the previous release:

source	rows with <ulcel>
pubtables1m	37%
ccocr	33%
omnidocbench16	25%
pubtabnet_icdar21	23%
tablet	22%
omnidocbench15	20%
ocrbenchv2	17%
scitsr	16%
fintabnet	14%

This release rebuilds the ground truth from the original HTML following PaddleOCR-VL 1.6's own parser (paddlex/inference/pipelines/paddleocr_vl/uilts.py::otsl_parse_texts):

• a cell starts at <fcel> (with text) or <ecel> (empty);
• a colspan puts <lcel> in the same row, to the right;
• a rowspan puts <ucel> in the row(s) below, in the same column;
• a cell merged both ways is <xcel>;
• rows are not padded — the row width is the row width.

Where a source's HTML in the previous release was only a shell (all cells empty, the text living in the source annotation), the source annotation is kept instead of the HTML-derived one, so no text is ever lost.

Verification. The canonical OTSL was parsed back with PaddleOCR-VL 1.6's own parser and compared cell by cell (text, rowspan, colspan) against the grid rebuilt from the HTML: 99.77% identical (2,993 / 3,000 samples; the 7 differences are extra empty cells produced by the official parser's internal "pad every row to max_cols" step, not structural errors).

Reference runs

scores_reference_*.parquet hold the per-sample scores of two reference runs against this ground truth, so the numbers below are reproducible:

run	model	pooled TEDS-S
base	PaddleOCR-VL 1.6 (no fine-tuning)	see scores_reference_base.parquet
sft_ckpt2800	our SFT (checkpoint-2800)	see scores_reference_sft_ckpt2800.parquet

The scorer is scripts/canonical_teds.py (primary metric TEDS-S; note that for pure-structure tables, whose cells are all empty, only TEDS-S is meaningful).

Intended use

Training and evaluating table-structure recognition models on 17 public and self-built table sources. The images are cropped table images; the labels are structure + cell text.

If you use this dataset, please also cite the original sources of the individual tables (the source column names them).