File size: 1,308 Bytes
a2827b8
251b517
9fe20a2
 
 
a2827b8
9fe20a2
a2827b8
9fe20a2
 
a2827b8
 
251b517
9fe20a2
2d64baa
9fe20a2
251b517
9fe20a2
 
 
 
 
 
 
 
 
 
 
2d64baa
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
---
title: Scientific Table Reconstruction
emoji: 📊
colorFrom: green
colorTo: yellow
sdk: static
app_file: index.html
pinned: false
license: mit
short_description: Scientific table reconstruction with RMS and exact mapping
---

# Scientific Table Reconstruction

A free static leaderboard comparing OCR pipelines and vision-language models on scientific table reconstruction, derived from TabulusBench Step 2.

Dataset: https://huggingface.co/datasets/SciKnowOrg/Scientific-Table-Reconstruction

The public test split contains 610 table images from 253 papers in five domains.
Each row contains only domain, paper_id, table_id, image and gold_rows.
Both RMS and Exact Mapping Match report macro and micro precision, recall and F1.
Evaluation mode: physical-table. Model input: image-only.

Open index.html via a local HTTP server to preview. The Space is entirely static.
See evaluation/README.md for the scoring protocol and contribution workflow.

The board starts without model entries. Sanity-check fixtures are never ranked.
New results are reviewed and committed by maintainers after scoring saved predictions.

The interactive comparison plotter supports model-size and score axes, type colors, domain filters, local summary previews and SVG downloads. Local previews do not enter public rankings.