---
license: other
task_categories:
  - image-to-text
language:
  - eng
  - zho
tags:
  - table-recognition
  - ocr
  - otsl
  - parquet
  - evaluation
---

# Table Recognition Datasets (Mixed)

Evaluation data for **table recognition** (OCR → table markup), packaged as
**parquet shards with the images embedded inline**.

Every table is stored in **two dialects** — OTSL and canonical HTML — so a model can be
scored against either without re-deriving one.

| Dataset | Rows | With image | Files | What it is |
|---|---|---|---|---|
| [`table_eval_paddleocr_vl`](tree/main/table_eval_paddleocr_vl) | 240,075 | 152,982 | 32 shards | The main evaluation set |
| [`table_eval_paddleocr_vl_excluded_toolong`](tree/main/table_eval_paddleocr_vl_excluded_toolong) | 3,828 | 3,633 | 4 shards | Samples withheld from evaluation because their tables are too long |

Start with the first. The second is an archive of exclusions, not a scoring target.

## Loading

```python
import pyarrow.parquet as pq

t = pq.read_table("table_eval_paddleocr_vl/shard_0000.parquet")
layout_only = t.filter(t.column("has_numeric") == False).to_pylist()
```

Images are the original encoded bytes, never re-encoded — recompressing would
introduce artefacts that contaminate the metric. When serving many requests, resolve
images lazily per request; do **not** call `pq.read_table(shard).to_pylist()` on an
inference path, since it re-decodes an entire row group per request and collapses under
concurrency.

## Scoring

Normalise prediction and ground truth to canonical HTML before computing TEDS, so a
model that answers in OTSL and one that answers in HTML are compared on equal terms.
The full column reference and distributions are in the
[main dataset card](tree/main/table_eval_paddleocr_vl).

For the SPRINT-MUSTARD family the ground truth is deliberately all-empty `<td>`, so
**only TEDS-S is meaningful** there; its TEDS column reads artificially low.