--- 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.