VERIDOC-QA: A Benchmark for Verifiable Multi-Document, Multi-Span Evidence Grounding

VERIDOC-QA (Verifiable Evidence across Documents for QA) is a question-answering benchmark for verifiable answer generation. It is the first QA benchmark to jointly require:

1. Multi-document synthesis. Answers must integrate information from up to three documents.
2. Multiple non-contiguous evidence spans. Supporting evidence is scattered across non-adjacent sentences.
3. Explicit span-level grounding. Every answer is validated against its evidence spans for both necessity and sufficiency via textual-entailment checks.

The benchmark targets trustworthy, evidence-backed QA: models must not only answer correctly but ground each answer in the specific sentences that support it.

Dataset Summary

VERIDOC-QA is built from MASH-QA (Zhu et al., 2020), an expert-curated extractive QA dataset in the consumer-health domain, through a multi-stage LLM-based pipeline:

1. Multi-document instance construction. Cross-document instances are created via multi-query fusion (merging related questions over distinct documents) and discourse segmentation (splitting one document into evidence-bearing segments).
2. Anaphora resolution. Pronouns and other anaphoric expressions in evidence sentences are resolved to explicit referents, making each span self-contained.
3. Answer generation & entailment-based validation. Abstractive answers are generated and accepted only if they are entailed by the full evidence set while no single span can be removed without breaking entailment, ensuring faithfulness and multi-span necessity.

Split	Instances	Description
single_doc	4,215	All evidence spans within one document
dual_doc	1,033	Evidence from two distinct documents (query fusion)
tri_doc	263	Evidence from three documents or segments (query fusion / document segmentation)
Total	5,511	Grounded in a corpus of 3,228 documents

On average, each instance requires 4.5 evidence spans, and both span count and answer length grow with document multiplicity.

Dataset Structure

The dataset has two configurations:

• qa (default): the QA instances, in three splits (single_doc, dual_doc, tri_doc);
• documents: the source corpus (split corpus), one row per document as a list of sentences.

qa fields

Field	Type	Description
id	int	Instance identifier, unique within each split
num_docs	int	Number of source documents (1, 2, or 3)
document_ids	list[int]	IDs of the source documents (keys into the documents config)
query	string	The question
evidence	list[string]	Evidence sentences as they appear in the source documents
evidence_after_resolution	list[string]	Evidence sentences after anaphora resolution (self-contained)
evidence_doc_ids	list[int]	For each evidence sentence, the ID of the document it comes from
evidence_sent_idx	list[int]	For each evidence sentence, its sentence index within that document
answer	string	Ground-truth abstractive answer, validated for faithfulness and span necessity

evidence, evidence_after_resolution, evidence_doc_ids, and evidence_sent_idx are parallel lists: position i in each refers to the same evidence span, and documents[evidence_doc_ids[i]][evidence_sent_idx[i]] == evidence[i].

documents fields

Field	Type	Description
document_id	int	Document identifier
sentences	list[string]	The document as an ordered list of sentences

Usage

from datasets import load_dataset

qa = load_dataset("ntunlplab/VERIDOC-QA", "qa")
docs = load_dataset("ntunlplab/VERIDOC-QA", "documents", split="corpus")

corpus = {d["document_id"]: d["sentences"] for d in docs}

ex = qa["tri_doc"][0]
print(ex["query"])
for doc_id, sent_idx, span in zip(ex["evidence_doc_ids"], ex["evidence_sent_idx"], ex["evidence"]):
    assert corpus[doc_id][sent_idx] == span
print(ex["answer"])

Supported Tasks

• Retrieval: retrieve the source documents for a query against the 3,228-document corpus;
• Retrieval-augmented generation (RAG): answer from top-k retrieved documents;
• Machine reading comprehension: answer given the gold documents;
• Evidence-grounded generation: generate an answer together with the supporting evidence spans.

Considerations

• Domain: all content is consumer-health text inherited from MASH-QA; findings may not transfer directly to other domains.
• Provenance: instances are anchored on expert-curated MASH-QA evidence spans; the multi-document structure, anaphora resolution, and answers are produced by an LLM pipeline with entailment-based validation and human spot-checks.
• Not medical advice: the dataset is intended for research on evidence-grounded QA, not for clinical use.

License

Released under CC BY-SA 4.0. Source text derives from MASH-QA.

Citation

The accompanying paper is currently under review. A citation entry will be added upon publication.