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.
|