---
title: FinLongDocQA Viewer
emoji: 📈
colorFrom: green
colorTo: blue
sdk: static
pinned: false
license: other
---

# 📈 FinLongDocQA Viewer

Static viewer for **FinLongDocQA** — a benchmark for *document-level numerical
reasoning across single and multiple tables in long financial annual reports*.

- Paper: <https://arxiv.org/abs/2604.03664>
- Code: <https://github.com/AI-Application-and-Integration-Lab/FinLongDocQA>
- QA dataset: <https://huggingface.co/datasets/Amian/FinLongDocQA>
- Report corpus (markdown, re-hosted): <https://huggingface.co/datasets/timchen0618/finlongdocqa-reports>

Two tabs:

- **📚 Corpus** — the 1,456 annual reports (`TICKER · YEAR`). A searchable
  dropdown selects a report (filter by ticker, fiscal year, or whether it has
  questions). Each report is a long 10-K converted to Markdown; the panel
  **renders it page-by-page** (page navigator + jump-to-page) and streams the
  markdown on demand from the HF dataset CDN. Page numbers match the evidence
  pages cited in the Eval tab.
- **❓ Eval** — the 7,527 numerical-reasoning questions. Each shows the question
  type (`table` / `text` / `mixed`), the ground-truth numeric answer, the
  **chain-of-thought reasoning trace**, the **executable Python** that computes
  the answer, and clickable **evidence-page chips** that jump straight to the
  cited page of the source report in the Corpus tab.

## Data

| File | Contents |
|---|---|
| `dataset_qa.jsonl` | 7,527 QA examples (source, from `Amian/FinLongDocQA`) |
| `eval.json` | 7,527 derived question records (Eval tab) |
| `corpus_index.json` | 1,456 derived report records (Corpus tab) |

The two derived files are regenerated with:

```bash
python scripts/build_data.py            # reads dataset_qa.jsonl + the reports/ tree
```

**The ~1 GB markdown corpus is not bundled.** The annual reports were downloaded
from the project's Google Drive, sanitized to valid UTF-8, and re-hosted on the
`timchen0618/finlongdocqa-reports` HF dataset. Each `corpus_index.json` record's
`md_url` points at that dataset's CDN, which serves markdown `inline` with
permissive CORS, so the viewer `fetch()`es and renders each report client-side
(via `marked`) without the Space hosting the corpus.

## Local dev

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

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

Dataset license: AI²Lab Source Code License (National Taiwan University).