- Who it helps
- Researchers and teams evaluating question answering over long PDFs.
- Availability
- The October 16 paper revision and public data repository are available. The release contains 13,548 training/development questions, rather than the full 15,000-question benchmark. Full assets use Git LFS, Git’s large-file download extension, and about 38 GB including cache; annotations can be read separately.
Inspect two different failures
The paper reports that supplying the correct supporting pages still leaves errors: its best reported accuracy in that setting is 81.7%. Better retrieval therefore cannot by itself establish correct visual interpretation.
The release has a clear boundary
The repository supplies annotations, PDFs and table workbooks, but no evaluation runner, judge implementation or model outputs. It states CC BY-SA 4.0 with source-specific exceptions. Reference answers must be kept out of the answering model’s inputs.
What you can do
- Read the released split table and evidence-field definitions before comparing a system against the paper. Open reference ↗
Where the evidence stops
Abstract and repository README checked. No dataset downloaded, document question answered or reported score reproduced.
Sources read. Sources checked October 21, 2026. No product workflow or benchmark was executed.
Sources & further reading
Paper · paper
VELD-QA: evidence grounding in long documents ↗
Read version 2 abstract and October 14/16 history. It distinguishes locating evidence from interpreting charts, tables and complex layouts.
VELD-QA authors · page checked 2026-10-21 · Published 2026-10-16
GitHub · implementation
VELD-QA released data and documentation ↗
Read the actual released train/development subset, approximately 38 GB checkout/cache requirement and CC BY-SA 4.0 with source-specific exceptions. Evaluation runner and model outputs are absent.
Oracle · page checked 2026-10-21
Coverage note: Independent hands-on testing and cross-platform discussion are incomplete. No missing links, creator identities or engagement counts are inferred.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
VELD-QA tests whether document assistants can locate the right visual evidence and correctly interpret charts, tables and complex layouts.