This dataset is a closed-world synthetic corpus for measuring knowledge injection via continued pre-training (CPT). It contains fictional materials, processes, facilities, and regulations rendered into documents (training) and multiple-choice questions (eval). All questions have deterministic ground truth derived from a hidden world-state database.

Repo: tensorvalley/cpt_qagen (public)
World ID: VYQ-24CS93V8
Generated: 2026-01-27 17:28:39 UTC
Novelty: Pronounceable pseudowords + unique markers + optional canaries
Evaluation: Closed-book MCQ, score by mean logprob per token

Document Examples (by style)

wiki-style

Pelbratau-37N is listed as Type-L08 in the Veylan cryometal index. It falls in
reactivity class RG-04 with melting band MB-28. Its certified process is Dunzorflux
(P-051). The signature tag is MK-4PEE.

manual-style

Material Sylgry-07Q carries taxonomy Type-L27. Core handling uses Dunbratulflux (P-065)
as the certified route. Facility restriction: Venzarmey-18S is banned.

compliance bulletin

Compliance note for Z-76Z: taxonomy Type-L57. Approved handling relies on Lumtekmarflux
(P-029). Restricted site: Marly-13R is disallowed for this material.

lab note

Lab note: Gormer-83M logged under class tier Type-L58. Process observed: Mersevflux (P-026).
Measured melting band MB-30; reactivity group RG-12 confirmed. Marker code recorded as MK-WGX6.

Question Examples (MCQ)

1-hop

Which process code is authorized for S-17A?
Options: ["P-000", "P-033", "P-063", "P-016"]  (answer_idx=1)

3-hop

Which flagship facility is named by the regulation tied to S-09T's approved process?
Options: ["Vexrissyl-60C", "Marly-13R", "Talvor-80U", "Sylsylven-81K"] (answer_idx=3)

Configs and Splits

This dataset uses multiple configs so that splits with different schemas can co-exist:

Config: `docs`

Split	Rows	Purpose
docs_train	61,360	CPT documents (no QA leakage)

Config: `qa`

Split	Rows	Purpose
qa_dev	10,000	Dev MCQ
qa_test	10,000	Test MCQ (distinct templates)

Config: `world` (optional)

Split	Rows	Purpose
world	20,680	Structured ground truth

Data Fields

docs_train

text: document text
meta: entity_id, entity_type, doc_type, world_id

qa_dev / qa_test

id: question id
prompt: MCQ prompt
options: list of answer options
answer_idx: correct option index
meta: type, hop, entity_id

world

type: entity type (material, process, facility, regulation)
id, name/title, attributes, relations, world_id

Methodology & Correctness Guarantees

This dataset is generated from a closed-world database (materials, processes, facilities, regulations) with deterministic, typed attributes and relations. Questions are rendered from that database, not inferred by an LLM.

Correctness is guaranteed by construction:

Each question’s answer is computed from the world DB (single-valued attributes or explicit relations).
Distractors are sampled from the same type domain (classification/process/facility/etc.).
Options are unique; the correct answer appears exactly once.
“None of the above” questions explicitly omit the correct answer and set the label accordingly.

You can verify correctness by re-deriving answers from world using entity_id and the question type/hop metadata.

Generation Process

World DB: Materials + attributes + relations to processes, facilities, and regulations.
Docs: Multiple paraphrased styles (wiki/manual/compliance/lab) with controlled noise.
QA: Template-disjoint MCQ with 1/2/3-hop reasoning + abstention checks.
Aliases: Documents mix canonical + secondary aliases; questions favor unseen alias.

Question Diversity

Hop depth mix (by default):

1-hop: 0.40
2-hop: 0.30
3-hop: 0.10
Compare/count: 0.10
Unanswerable: 0.10

Families (examples): classification, signature marker, approved process, process→regulation, facility→regulation, prime facility via 3-hop composition, shared classification, process count bins, and abstention.

Phrasal / Template Diversity

Docs: 4 styles (wiki, manual, compliance, lab) with shuffled sentence order and synonym substitutions (classification/reactivity/melting/approved/banned/signature).
QA: Dev and test use disjoint template sets to avoid template memorization.
Noise: Optional filler sentences at a configurable rate.
Aliasing: Documents prefer canonical + one alias; questions prefer a different alias.

Generation Settings

Materials: 20,000
Processes: 400
Facilities: 200
Regulations: 80
Classifications: 64
Melting bands: 32
Reactivity groups: 16
Docs per material/process/facility/regulation: 3/2/2/2
Noise rate: 0.1
Canary rate: 0.001
QA mix (1/2/3-hop, compare, unanswerable): 0.40/0.30/0.10/0.10/0.10

Intended Use

Measure knowledge injection from CPT without retrieval or judge models.
Control difficulty via hop depth, attribute cardinality, and redundancy.
Sanity-check novelty with canary strings and pre-CPT chance performance.

Limitations

Synthetic language and taxonomy; not a proxy for real-world discourse.
Alias coverage is systematic but limited to a small set per material.
Designed for closed-book MCQ evaluation; not for open-ended QA.

Citation

If you use this dataset internally, cite the repository tensorvalley/cpt_qagen and the world id VYQ-24CS93V8 in your experiment logs.

Downloads last month: 25

You need to agree to share your contact information to access this dataset

Dataset Summary