deepinquiry/verified-facts-sample-100

DeepInquiry Verified Facts (Sample-100)

A 90-fact sample from the DeepInquiry verified-facts corpus. Every fact in this sample has been cross-checked against multiple structurally independent web sources, cited, dated, and confidence-scored before it entered the corpus.

This is a preview sample. The full corpus (~942 approved facts as of Sept 2026, growing continuously) is available via the DeepInquiry API at deepinquiry.ai/pricing and — pending qualification — via AWS Data Exchange.

TL;DR

• —90 verified facts across 8 categories
• —Gold and Silver verification tiers only
• —Median 3 independent source clusters per fact
• —Every fact has an independence score (0.0–1.0) that penalizes wire-service duplicates and Wikipedia mirrors
• —One fact per line in JSONL, ready to load with datasets.load_dataset("deepinquiry/verified-facts-sample-100") or pandas.read_json(lines=True)

Dataset structure

Files

Fields

Sample record

{
  "id": 31,
  "fact_text": "Tardigrades can survive in the vacuum of space, withstand pressures six times greater than the deepest ocean trenches, and endure radiation doses 1,000 times what would kill a human.",
  "category": "nature",
  "era": "modern",
  "region": "global",
  "verification_tier": "silver",
  "independence_score": 1.000,
  "cluster_count": 7,
  "surprise_score": 9,
  "confidence_level": "high"
}

The full DeepInquiry API also returns a sources array on every response (URL, publisher, cluster ID, domain class). To keep this sample compact and portable, the sources array is omitted here — see the full API for the complete provenance trail per fact.

Loading the dataset

from datasets import load_dataset

ds = load_dataset("deepinquiry/verified-facts-sample-100")
print(ds["train"][0])

Or without datasets:

import pandas as pd
df = pd.read_json("data/train.jsonl", lines=True)

Verification methodology (short version)

Every fact in the DeepInquiry corpus goes through this pipeline before it becomes queryable:

1. 1.Candidate extraction. Claims are extracted from open-web sources across multiple domains.
2. 2.Multi-source corroboration. The pipeline searches for independent corroboration across the open web. A claim needs at least two structurally independent sources to advance.
3. 3.Independence scoring. Sources are clustered such that wire-service duplicates, syndicated re-posts, and Wikipedia mirrors are collapsed into a single cluster. The independence score reflects how many distinct clusters back the claim — not how many URLs happen to mention it.
4. 4.Verification tier. Facts are graded gold, silver, or bronze based on evidence strength (cluster count, source authority, cross-domain agreement).
5. 5.Citogenesis check. Facts are flagged when the source cluster shows signs of copying from a single seed (the "citation loop" problem in AI-generated content).
6. 6.Human-in-the-loop review. Approved facts enter the queryable corpus.

Wikipedia is cited when relevant but is excluded from the independence calculation — because Wikipedia mirrors are not independent sources of a claim.

Distribution

Intended uses

• —RAG grounding. Use these facts as a curated upstream truth source for retrieval-augmented generation, where you want your LLM's outputs backed by evidence with a documented independence signal.
• —Fact-checking research. Study how independence-scoring and cluster-count metadata affect downstream classifier or verifier training.
• —Trivia and quiz apps. 90 facts across broad categories with a surprise score for user-facing curation.
• —Edtech. Reference content where wrong facts have real user cost.
• —Evaluation. Benchmark your own knowledge base or verification pipeline against a corpus that publishes its methodology.

Out-of-scope uses

• —Not a legal, medical, or financial reference. These facts are verified as of the date they entered the corpus. The world changes; downstream applications for high-stakes decisions must validate against their own primary sources.
• —Not a real-time news feed. DeepInquiry is a reference corpus, not a news API.
• —Not a training-data substitute for LLMs. This is a sample. Training-scale use requires the full API and a data-license conversation.
• —No PII, no sensitive-category data. The corpus is publicly-known facts about the world, not information about private individuals.

Limitations and biases

Small sample. 90 facts is not statistically representative of the ~942-fact full corpus and definitely not of "all human knowledge". Do not draw strong conclusions about coverage, geographic balance, or category balance from this sample alone.

Category skew in the full corpus. As of Sept 2026 the full DeepInquiry corpus over-indexes on science (~55% of approved facts), history (~26%), and nature (~6%). Other categories are actively growing. This 90-fact sample rebalances to 8 categories with ~12–15 facts each, so it is intentionally more balanced than the underlying corpus. Downstream users who assume the full corpus has this balance will be surprised.

English-only. All facts in this sample and in the current DeepInquiry corpus are English-language. Multilingual coverage is on the roadmap.

Western-source bias. Independent web sources still skew heavily toward US and Western European publishers. A fact scoring 1.0 on independence still probably means "corroborated across multiple Western newsrooms" not "corroborated across globally distributed sources." We are actively broadening the source pool but this bias is real today.

Verification is not omniscience. A gold-tier fact is a fact where multiple independent sources agreed at the time of verification. It is not a guarantee of factual correctness — the sources themselves may have been wrong.

Snapshot in time. This sample was extracted from the corpus on 2026-09-01. Facts in the live corpus may have been updated, re-verified, or deprecated since.

Ethical considerations

• —No sensitive personal data. The corpus contains no PII. All people mentioned are public figures or historical persons.
• —Sources are cited on every API response. In this sample the sources field is omitted for compactness, but the full DeepInquiry API always returns the source cluster. If you build a product on this sample and later want the full sourcing, upgrade to the API.
• —Independence scoring is a signal, not a truth-o-meter. A fact with independence_score = 1.0 is a fact with high structural source diversity. It is not a claim that we've solved epistemology.

License

CC BY 4.0 — Creative Commons Attribution 4.0 International.

You may use, redistribute, remix, and build on this dataset for any purpose (including commercial), as long as you provide attribution:

DeepInquiry Verified Facts (Sample-100), Epic Sky LLC, licensed under CC BY 4.0. Source: https://huggingface.co/datasets/deepinquiry/verified-facts-sample-100

The DeepInquiry API and full corpus are commercially licensed separately — see deepinquiry.ai/pricing.

Full API

If you want:

• —The full ~942-fact corpus (growing continuously)
• —The complete sources array with URLs, publisher, cluster ID, and domain class on every fact
• —Bronze-tier facts in addition to Gold and Silver
• —Search, filter, and per-fact lookup endpoints
• —Audit-mode fields for compliance and governance workflows

→ Subscribe at deepinquiry.ai/pricing — Free tier is 500 requests / month with silver+ facts. Paid tiers start at $29 / month.

Coming soon on AWS Data Exchange.

Citation

If you use this dataset in research or a product, please cite:

@misc{deepinquiry_sample_100,
  author = {Rush, Bryan and {Epic Sky LLC}},
  title = {DeepInquiry Verified Facts (Sample-100)},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/deepinquiry/verified-facts-sample-100}},
  note = {A verified-facts sample from the DeepInquiry corpus, licensed under CC BY 4.0}
}

Contact

• —Web: deepinquiry.ai
• —Publisher: Epic Sky LLC
• —Hugging Face issues: open a discussion in the Community tab
• —API support: hello@deepinquiry.ai