CheneyWong/real-scene-hard-span-1600

Real-scene Hard Tables with Spans — 1,600 samples

This dataset contains 1,600 table crops from 1,600 distinct original images, together with the original image bytes and existing OvisOCR3 and MinerU predictions. Source: CheneyWong/real-scene-tables, revision 7eb6a7cd173bd14a833382aea9fa824647cdfcb1.

These are model predictions, not human ground-truth annotations. Difficulty indicates disagreement among model predictions; it does not establish which prediction is correct.

Selection

1. Tables were detected and cropped using PPDocLayout-v3; the crop is lossless PNG at the original resolution, using the saved bounding box after EXIF orientation.
2. Both exported predictions, OvisOCR3 and MinerU, must parse to at least one merged cell with rowspan > 1 or colspan > 1. HTML and OTSL are supported. A CSS <span> tag or a unit span does not qualify.
3. A sample is hard when zero or one of six fixed one-way model comparisons has TEDS-S >= 0.95. The six (anchor, candidate) directions are:
   • PaddleOCR-VL-1.6 → OvisOCR3
   • PaddleOCR-VL-1.6 → MinerU-4.1.3
   • PaddleOCR-VL-1.6 → HunyuanOCR-1.5
   • OvisOCR3 → MinerU-4.1.3
   • OvisOCR3 → HunyuanOCR-1.5
   • MinerU-4.1.3 → HunyuanOCR-1.5
4. Selected from 12,688 hard samples satisfying the two-model span condition, with fixed seed 20261102, retaining distinct parent images. No inference or scoring was repeated for this export.

There are 1,142 samples with zero qualifying comparisons and 458 with one. Truncated/error predictions and incomplete scores were excluded from hard classification.

Fields

Field	Meaning
sample_id, image_id	Stable table and parent-image identities
original_bytes	Exact original encoded image bytes; binary, not a file path
crop_bytes	Lossless cropped table PNG bytes; binary, not a file path
ovisocr3_result, mineru_result	Existing prediction text used by the scorer; may be HTML or OTSL
ovisocr3_raw_result, mineru_raw_result	Preserved raw model output where available
ovisocr3_has_span, mineru_has_span	Both are true in this export
difficulty, good_pair_count	hard, and the number of selected TEDS-S values >= 0.95
teds_s_selected_pairs	Six saved scores, in the direction order listed above
bbox_xyxy, polygon_xy	Crop coordinates and detector polygon in the EXIF-oriented original image frame
original_sha256, crop_sha256	SHA-256 of the exact encoded binary image values
source_shard, source_row, source_url	Source provenance; row indices are zero-based
dataset_repository, dataset_revision	Source repository and pinned revision
image_width, image_height, crop_width, crop_height	Image dimensions
qa_status, qa_flags	Retained crop quality diagnostics
scene_type, language, rule_version	Source metadata and classification rule version

Quality checks and limitations

All 1,600 exported original/crop byte hashes were verified after writing the Parquet. Each decoded crop was checked pixel-for-pixel against the saved bounding box of the EXIF-oriented original. Both exported model predictions were checked for parsed merged cells.

These checks establish byte identity and geometric correspondence; they do not establish that detection captured every table cell or that either prediction is correct. 714 samples have qa_status=auto_pass; 886 retain qa_status=needs_review. The latter are included with their original diagnostics and have not all been manually reviewed. auto_pass is an automated status, not a human annotation.

The Parquet file contains encoded bytes directly; image decoding should use Pillow, for example:

import io
import pyarrow.parquet as pq
from PIL import Image, ImageOps

row = next(pq.ParquetFile("train-00000-of-00001.parquet").iter_batches(batch_size=1)).to_pylist()[0]
original = ImageOps.exif_transpose(Image.open(io.BytesIO(row["original_bytes"])))
crop = Image.open(io.BytesIO(row["crop_bytes"]))
print(row["ovisocr3_result"])
print(row["mineru_result"])

Original image rights remain with their respective holders. This export does not grant additional rights over source images.