David-beakr/wiki-ingest-health-seed

Internal Beakr dataset — not for public distribution. Host as a private repository.

Wiki Ingest Health Seed v0.4.1

A dataset for Pillar 5: wiki / knowledge-layer health. The pillar question is broader than "did ingest extract the right things": after ingesting a tiny corpus, did the wiki preserve the important facts, cite them properly, organize them reasonably, and make them retrievable — without inventing what isn't there, and without producing duplicates on re-ingest?

Stats

Entities	9
Facts	13
Relationships	9
Distractors	4
Idempotence tests	1
Canned questions	3
Health dimensions	6
Check types	7
Corpus	6 documents, ~670 words, shared with kg-retrieval-seed
Schema version	1.0
License	Internal-only (synthetic corpus)

What's in this repo

wiki-ingest-health-seed/
  README.md
  gold.json    expected entities, facts, relationships, distractors, idempotence, canned questions
  corpus/      6 fixture documents (~670 words)

The corpus is shared with kg-retrieval-seed: this dataset tests whether ingest produced a healthy wiki; that one tests retrieval/QA over the same docs. Keep both in sync if the corpus changes.

Load:

from huggingface_hub import snapshot_download
import json
path = snapshot_download(repo_id="David-beakr/wiki-ingest-health-seed", repo_type="dataset")
gold = json.load(open(f"{path}/gold.json"))

The six health dimensions

Dimension	Definition	How tested	Checks
Coverage	Expected entities, decisions, owners, dates, topics are represented	9 entities + 13 facts must appear	page_type_correct (9), fact_in_wiki (13)
Factual accuracy	Gold facts preserved without invention	Facts present with correct values; distractor claims defended	fact_in_wiki (13), fact_not_asserted (5 on distractors)
Citation / provenance	Important claims point back to the right source	Each fact carries expected_citation_tokens; canned questions require citation	expected_citation_tokens on fact_in_wiki, qa_correct_with_citations
Organization	Content lands in the expected type and avoids duplicates	Pages have the right Beakr type; relationships exist as edges; re-ingest doesn't duplicate	page_type_correct (9), relationship_in_graph (9), no_duplicate_pages
Retrieval usefulness	Canned questions answerable with the right facts and citations	3 canned questions, hybrid scoring	qa_correct_with_citations (3), llm_judge_rubric (3)
Idempotence	Re-ingesting the same content creates no duplicate or contradictory content	One re-ingest test	no_duplicate_pages (1)

What the dataset tests

• 9 entities — six Beakr page types exercised: initiative (Project Harbor), person (Maria Chen, Devin Park, Aisha Okafor), organization (InfoSec team, Legal team), topic (Meridian), decision (platform selection), meeting (Jan 26 decision sync). The other six page types (overview, research_note, experiment, protocol, compound, dataset) are out of scope — the corpus contains no content of those types.
• 13 facts — project-level facts (owner, target date, platform, scope, budget, milestones, blockers), each with a canonical claim_key (e.g. project_harbor.owner), an expected_value, and an expected_citation_tokens set.
• 9 relationships — graph edges: owner-of, due-date, action items, blockers, decision-to-project links.
• 4 distractors — specific false claims that must NOT be asserted; they share a source (DOC-05, a superseded draft) and contradict specific facts.
• 1 idempotence test — re-ingest of identical content must not produce duplicate pages.
• 3 canned questions — natural-language questions over the wiki, hybrid-scored (substring + judge).

Gold shape

{
  "expected_entities":      [{ "id", "type", "name", "aliases", "checks" }],
  "expected_facts":         [{ "id", "claim", "source_docs", "notes", "checks" }],
  "expected_relationships": [{ "id", "derived_from_facts", "source_docs", "notes", "checks" }],
  "known_distractors":      [{ "id", "false_claim", "contradicts_fact", "true_value", "distractor_source", "checks" }],
  "idempotence":            { "id", "designated_doc", "notes", "checks" },
  "canned_questions":       [{ "id", "question", "expected_answer", "answer_facts", "retrieval_challenge", "notes", "checks" }]
}

Check types

• page_type_correct — the entity is assigned the correct Beakr page type. Params: entity_name, expected_type (one of Beakr's 12 page types). Deterministic.
• fact_in_wiki — claim_key resolves to expected_value on the expected page; optional expected_citation_tokens for provenance. Deterministic.
• fact_not_asserted — claim_key does NOT resolve to a forbidden value (must_not_equal / must_not_contain). Deterministic.
• relationship_in_graph — the graph contains a subject-predicate-object edge. Deterministic.
• no_duplicate_pages — re-ingest of identical content produces no net new pages or facts. Deterministic.
• qa_correct_with_citations — the generated answer contains required substrings, asserts no forbidden claims, and cites expected sources. Deterministic.
• llm_judge_rubric — the generated answer satisfies a per-question rubric. Judge.

Principle: deterministic checks where wiki output is structured (a typed page with a claim_key:value body, edges to other pages, source citations — verifiable by key lookup or graph query); judge checks where the wiki is exercised through a generated answer. The 3 canned questions are hybrid (both layers hard) because substring matching can pass an answer that includes the right keywords without answering — e.g. "Maria Chen owns Project Harbor but the target completion is unclear" passes keyword checks while failing Q-01. The judge catches that. This is the closed-form-vs-judge scoring policy applied across the Beakr pillar set (cited rationale in the Long-Term Memory and Scientific Work seeds): deterministic gating only on closed-form properties; semantic properties owned by a tight-rubric judge so a brittle lexical proxy never hard-vetoes a correct paraphrase or negation. The deterministic checks here verify structured wiki output (typed pages, claim_key:value, graph edges) — inherently closed-form — so no check changed.

Design grounding

No single academic paper evaluates exactly what Beakr's knowledge layer does. Two extraction-focused anchors plus Beakr's own health docs:

• iText2KG (Lairgi et al., WISE 2024; arXiv:2409.03284) — incremental, zero-shot KG construction (Documents Distiller, Incremental Entities/Relations Extractors, Graph Integrator); the closest method-level analog to Beakr's compiler agent.
• Text2KGBench (Mihindukulasooriya et al., ISWC 2023; arXiv:2308.02357) — KG-extraction benchmark; seven metrics across fact extraction (precision/recall/F1), ontology conformance, and hallucination.
• Beakr's graph-health signals (docs.thebeakr.com/platform/graph) — backlink count, orphan count, red-link density, thin content, staleness, source agreement. The more complete health anchor; not yet tested here (see limitations).

Anchor fit is partial: iText2KG outputs entity-relation triples while Beakr outputs typed pages with type-specific fields; Text2KGBench evaluates against a fixed ontology while Beakr's wiki involves [[link]] parsing, temporal metadata, title aliases, red links, and source-join records. Extraction metrics cover ~3 of the 6 health dimensions cleanly (Coverage, Factual accuracy, Citation quality), partially cover Organization; Retrieval usefulness and Idempotence are addressed via Beakr-specific mechanisms.

Concept mapping:

This dataset	iText2KG	Text2KGBench
expected_entities + page_type_correct	Incremental Entities Extractor	(no analog — ontology classes, not typed pages)
expected_facts + fact_in_wiki	(no analog — triples, not typed facts)	Fact extraction (P/R/F1)
expected_relationships + relationship_in_graph	Incremental Relations Extractor	Relation extraction
known_distractors + fact_not_asserted	(no analog)	Object hallucination
idempotence + no_duplicate_pages	Graph Integrator dedup	(no analog)
canned_questions + QA-with-citations	(no analog — downstream task)	(extrinsic evaluation)

Methodology deviations

• Scale. ~670 words across 6 documents, 13 facts, vs iText2KG's many thousands of sentences and Text2KGBench's 13,474 + 4,860. Intentional for a seed.
• Typed pages, not ontology triples. page_type_correct tests whether the entity became the right kind of page, not whether a triple used the right predicate.
• Hybrid scoring on canned QA. The 3 canned questions pair a substring check with an llm_judge_rubric hard check.

Harness requirements

No external fetching — the corpus ships in corpus/. Ingest the corpus into the wiki under test, then: assert entities exist with the right page types; assert facts present with correct citations; assert relationships in the graph; assert distractor facts NOT asserted; re-ingest the designated idempotence doc and verify no duplicates; run the 3 canned questions and grade. The structural checks require harness access to wiki/graph introspection (page types, claim_key resolution, graph edges, re-ingest dedup); the 3 canned questions require an LLM judge.

Known limitations

Beakr graph-health signals not yet tested (documented at docs.thebeakr.com; each would need a new check type):

• Orphan detection — no check asserts every entity has at least one inbound edge.
• Backlink count quality — inbound link counts per page are not measured.
• Red-link density — unresolved [[link]] references are not tested.
• Thin content — pages below a minimum content threshold are not flagged.
• Staleness — pages past an update threshold are not tested.
• Source agreement — generalized confirm/contradict scoring is not implemented (only the 4 specific distractors).

Extraction-quality gaps:

• Page-type coverage is partial — 6 of Beakr's 12 page types are exercised; the other 6 (overview, research_note, experiment, protocol, compound, dataset) need corpus content of those types. The life-sciences types (experiment, protocol, compound, dataset) are the highest-value additions.
• Temporal metadata not tested — event_start / event_end / date_precision extraction is unverified.
• Title aliases not tested — entities list aliases, but preservation through ingest is unverified.
• [[link]] parsing not tested as a distinct mechanism (edges are checked, not their bracket-link origin).

Evaluation-scope gaps:

• General hallucination beyond the 4 specific distractors is not caught.
• Distractor coverage is thin — 4 distractors across 13 facts.
• No multi-source agreement testing — DOC-03 and DOC-06 contain overlapping facts; resolving agreement vs disagreement is not tested.

Designed declaratively before harness integration; the page_type_correct mechanism and check parameters may shift once integration begins.

Out of scope

• Broad semantic quality ("is this page well-written").
• Long-horizon wiki drift (weeks of edits).
• Human editorial usefulness.
• Connector-specific behavior — source tokens are treated uniformly regardless of provider.
• Multi-turn ingest.
• Adversarial input beyond the 4 distractors (e.g. prompt injection inside source documents).
• Cross-organization isolation — single-tenant fixtures only.

License & attribution

Internal-only Beakr compilation; the corpus is synthetic. Anchors, cited not redistributed: iText2KG (Lairgi et al., WISE 2024, arXiv:2409.03284); Text2KGBench (Mihindukulasooriya et al., ISWC 2023, arXiv:2308.02357).

Changelog

• 0.4.1 — Robustness follow-up scan (AUDIT §9), closed-form-vs-judge policy (now cross-referenced under Check types). DIS-04: re-keyed the scope must_not_contain from "archived" to "all customer records" — a correct, complete scope value ("active records only; archived accounts excluded") contains "archived", so the old token was a negation collision; the distractor's distinctive quantifier is the robust discriminator. Q-01: narrowed the must_not_assert distractor tokens from bare "Tom Reyes"/"February 28" to present-tense assertions ("owned by Tom Reyes", "due February 28", …) so a correct answer that refutes the distractor (or retains it as superseded history, which DIS-01 expects) isn't rejected. No entities, facts, relationships, or tested properties changed.
• 0.4.0 — Conformed to the canonical schema: normalized the singular check to a one-element checks[] on entities, facts, relationships, and the idempotence test; schema_version → 1.0. The 6-array shape is retained (a dedicated adapter maps it to EvalCase at load). No gold content changed.
• 0.3.2 — Removed the stale "Legacy reference" note (the companion's gold no longer carries the old name); tightened the README.
• 0.3.1 — Renamed wiki-ingest-seed → wiki-ingest-health-seed; reframed the README around the six health dimensions. Gold content unchanged from 0.3.0.
• 0.3.0 — Added page_type_correct checks on all 9 entities; hybrid scoring on the 3 canned questions; grounded in iText2KG + Text2KGBench.
• 0.2.1 — (under the predecessor name wiki-ingest-seed) added DOC-06 with the staging-environment and encoding-bug facts and relationship (FACT-12, FACT-13, REL-09).