QinghaiRAG dataset card

QinghaiRAG packages provenance-aware sources, entities, verified facts, release-safe chunks, and QA evaluation records for Qinghai intangible cultural heritage and regional culture.

Current status

Version 1.0.0-rc1 is a real-data benchmark candidate, built from registered official sources and reviewed local-government attachments. It is no longer the original toy engineering seed.

The latest audited build contains:

Table or review layer	Records
Registered sources	868
Canonical entities	653
Verified facts	1,595
Release-safe chunks	948
QA records	1,000
Manually checked facts	511
Manually checked QA	300

This candidate meets the configured V1 ranges for sources, QA, manually checked facts, and manually checked QA. It remains below the repository's final V1 targets for entities (2,000), facts (10,000), and chunks (10,000), so it must be described as a release candidate rather than the final benchmark-grade release.

Configurations

The Hub repository uses configs/subsets because the tables have different schemas. Each config currently exposes a full split:

• source_registry: source and licence registry.
• entities: canonical entities and aliases.
• facts: evidence-backed triples.
• open_chunks: full-text-compatible or project-authored synthetic retrieval units.
• qa_benchmark: answerable and unanswerable evaluation questions.
• audit_samples: de-identified accepted review samples and correction records.

For example:

from datasets import load_dataset

facts = load_dataset("zhangjh123/QinghaiRAG", "facts", split="full")
qa = load_dataset("zhangjh123/QinghaiRAG", "qa_benchmark", split="full")
audits = load_dataset("zhangjh123/QinghaiRAG", "audit_samples", split="full")

The QA config intentionally remains a single full split in this release candidate. A future train/validation/test layout will use entity- and source-aware grouping rather than random row splitting, which would leak near-identical facts across partitions.

Intended uses

Dataset engineering research, citation-aware retrieval prototypes, Graph RAG interfaces, evidence-retrieval comparisons, refusal evaluation, and education about provenance/licensing.

Out-of-scope uses

Authoritative heritage records, legal or policy advice, commercial travel recommendations, replacing communities or heritage bearers, reconstructing restricted source text, identifying private individuals, or treating generated QA as an independent measure of world knowledge.

Provenance and review

Every released fact and answerable QA record resolves to registered source_id and source_url fields. Current source families include the China Intangible Cultural Heritage website, Qinghai provincial cultural-tourism pages, county-government publications, reviewed official attachments, and the Qinghai Tibetan Culture Museum official API.

verified=true means a record passed the pipeline's source and consistency checks. manual_checked=true is stricter: a reviewer inspected the exact evidence snapshot. QA approvals are hash-bound, so a regenerated or edited question cannot inherit an earlier manual decision.

The release validator reports zero restricted open-text violations. Open text is included only where the source-level release policy permits it; other sources contribute metadata, reviewed facts, or project-authored synthetic factual summaries.

A pre-publication evidence-minimization audit rewrote 126 legacy table excerpts—62 category facts and 64 applicant facts—to retain only the project name and relation-specific value. Fact IDs, triples, source URLs, and review states were unchanged. The validator rejects future evidence excerpts containing unrelated sex, birth-date, address, telephone, or identification fields.

Retrieval baselines

The audited 1,000-question run uses top_k=5. These are evidence-retrieval metrics, not generative-answer correctness scores.

Mode	Recall@5	MRR	Unanswerable refusal	Citation presence
vector-only	0.9767	0.9362	1.0000	1.0000
bm25-only	0.9922	0.9745	1.0000	1.0000
graph-only	0.9978	0.9889	1.0000	1.0000
hybrid	0.9967	0.9881	1.0000	1.0000
vector-only+rerank	0.9767	0.9579	1.0000	1.0000
bm25-only+rerank	0.9922	0.9738	1.0000	1.0000
graph-only+rerank	0.9978	0.9709	1.0000	1.0000
hybrid+rerank	1.0000	0.9787	1.0000	1.0000

Answerable QA is generated from verified facts and then sampled for manual review. Consequently, these scores measure in-dataset evidence retrieval and should not be presented as performance on unseen external knowledge. Refusal uses a conservative requested-attribute support check: retrieval of a related entity alone is not treated as evidence for an absent field.

Limitations

Coverage is incomplete and geographically imbalanced, and source availability can change. Official publication does not eliminate errors, historical terminology differences, or institutional viewpoint bias. Entity aliases and short personal names can be ambiguous. Most facts are verified automatically rather than manually checked. Text-only release omits multimodal, spatial, performative, and oral dimensions central to many heritage practices.

Licensing

Code licensing does not determine dataset licensing, and the dataset repository has no blanket content licence. Inspect each source_registry record for license_status, release_policy, raw_text_release, attribution requirements, and the original URL. Raw/local-only and restricted content is excluded. See docs/SOURCE_POLICY.md.