README.md

metadata

title: MuDABench Viewer
emoji: 📊
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
license: apache-2.0

📊 MuDABench Viewer

Static viewer for MuDABench — a benchmark for multi-document analytical question answering over large-scale financial document collections (Chinese A-share + US market filings).

• Paper: https://arxiv.org/abs/2604.22239 (ACL 2026 Findings)
• Code: https://github.com/Zhanli-Li/MuDABench
• Dataset: https://huggingface.co/datasets/Zhanli-Li/MuDABench

Two tabs:

• 📚 Corpus — every source document (589 PDFs). A searchable dropdown selects a document by title (symbol · year · doctype); filter by doc type or year. The panel shows the document's structured metadata — every value_* field observed across the questions that cite it, with its schema description — and renders the PDF inline (streamed from the Hugging Face dataset CDN). A collapsible list links to each question that references the document.
• ❓ Eval — the 332 analytical questions (166 simple + 166 complex; toggle between the two sets). Each question shows the gold final answer, the supporting facts (source_answer), and the supporting documents (each value_* field + description; click a document to jump to it in the Corpus tab).

Data

File	Contents
simple.json	166 questions with concise final answers (source, from HF)
complex.json	166 questions with longer analytical final answers (source, from HF)
eval.json	332 derived question records (Eval tab)
corpus_index.json	589 derived document records (Corpus tab)

The two derived files are regenerated from the source JSON with:

python scripts/build_data.py

PDFs are not bundled (the corpus is ~4 GB). Each corpus record's pdf field points at the dataset CDN (https://huggingface.co/datasets/Zhanli-Li/MuDABench/resolve/main/data/pdf/<id>.pdf), which serves PDFs inline with permissive CORS, so the <iframe> renders them directly without the Space having to host them.

Local dev

python -m http.server 8000   # then open http://localhost:8000/

Serves statically — no build step. Push to the HF Space remote to deploy.

Dataset & code license: Apache-2.0.