LoreHopQA measures whether a retrieval-augmented system uses the context it was given. That measurement holds only while no evaluated model has encountered the collection during training, which is why access is gated and approved manually by the authors. Access is granted under the LoreHopQA Data Use Agreement, distributed with the collection as LICENSE. Please read it, fill in the form and accept each term below. Evaluation and publication are unrestricted; training on the items is not.

🌍🔎 LoreHopQA

A multi-hop question-answering collection set in 61 invented worlds, built so that the retrieved context is the only available source of the answer. Every entity, institution, event and place name is coined, so an answer cannot be recalled from pre-training.

🚦 Access Conditions

This collection is gated, and access is approved manually by the authors.

To request access, log in, read the LoreHopQA Data Use Agreement, then fill in the form on this page and tick every term of use. Requests that do not accept all terms are not approved. You will be notified by email.

Accepting the form is acceptance of the agreement. In short, and without replacing it:

	Items	Lorepedia
Evaluate models	✅	✅
Publish metrics and analyses	✅	✅
Train or fine-tune a retrieval model	❌	✅ under Section 5
Train or fine-tune a generative model	❌	❌
Redistribute	❌	❌
Send to a retaining or training API	❌	❌

📝 Dataset Summary

Standard multi-hop benchmarks are built on Wikipedia, so a model can often answer from parametric memory and an end-to-end score cannot separate retrieval quality from memorisation. LoreHopQA removes that route: the questions concern entities that do not exist, so the supplied documents are the only place the answer can be found.

20,612 items, 16,490 test and 4,122 validation. No training split is released. Items follow the HotpotQA schema, so the official scorer runs unmodified.

Language: English
Paper: LoreHopQA: Decoupling Answerability from Parametric Knowledge in Multi-Hop Retrieval-Augmented Generation (under review)
Authors: Juan Filgueiras, Marcos Fernández-Pichel, Javier Parapar (IRLab, CITIC, Universidade da Coruña)
Built from: structural transposition of HotpotQA evidence layouts, and instruction induction
Code: the validation experiments are at IRLab-UDC/lorehop-qa

🎯 Uses

Direct Use

Measuring whether a RAG system uses its retrieved context, with no parametric floor.
Studying ranking position, context redundancy, citation accuracy and evidence layout in isolation from pre-training exposure.
Comparing first-stage retrievers where domain familiarity is a variable you set rather than one inherited from a training mixture.

Out-of-Scope Use

Training or fine-tuning any model on the items. This is prohibited and is what the collection's value depends on.
Redistribution, or upload to any hub, harness or leaderboard that exposes item content.
Treating the invented worlds as factual. They are fiction, written to be mundane.

📦 Dataset Structure

Path	Contents
`scorer/test.jsonl`	Test split in HotpotQA's own field layout, for the official scorer
`scorer/validation.jsonl`	Validation split, same layout
`data/test.jsonl`, `data/validation.jsonl`	The same items as parallel lists, for `load_dataset`
`lorepedia/lorepedia.jsonl`	97,411 background articles, relevant to no question
`examples/`	Ten illustrative items in a world of their own, with a rendered walkthrough
`statistics.md`	Composition by construction path, question family, difficulty and split
`LICENSE`	The LoreHopQA Data Use Agreement

The release carries two encodings of the same items, because the scorer and load_dataset need different ones. scorer/ keeps HotpotQA's context as [title, [sentences]] pairs, so the official scorer runs unmodified. Arrow cannot type that shape, so data/ carries the same records with context as {"title": [...], "sentences": [[...]]}, which load_dataset reads directly. Neither is a sample of the other.

Fields

_id, question, answer, type, level, context and supporting_facts follow HotpotQA. type additionally takes chain and convergent, which the scorer ignores.

_id is opaque and carries no information beyond the construction path: trp- for a transposed item, nov- for an induced one.

The ten items under examples/ are not part of any split and must not be scored against. They exist so the format can be read in full without publishing an item a model might later be evaluated on.

📊 Dataset Statistics


Items (test / validation)	20,612 (16,490 / 4,122)
Invented worlds	61
Construction path (transposed / induced)	7,371 / 13,241
Question family (bridge / comparison / chain / convergent)	5,892 / 5,046 / 7,192 / 2,482
Difficulty (easy / medium / hard)	1,324 / 5,611 / 13,677
Candidate documents per item (median, range)	6, 4-10
Relevant documents per item (median, range)	2, 2-4
Supporting facts per item (median)	3
Question / answer length in words (median)	20 / 2
Yes/no answers	11.9%
Item documents	118,794
Lorepedia articles / words	97,411 / 8.01M

Per-split and per-subset figures are in statistics.md.

🚀 Usage

from datasets import load_dataset

# Gated: log in first with `hf auth login`, or pass token=... to load_dataset
qa = load_dataset("irlab-udc/lorehopqa")                      # test + validation
corpus = load_dataset("irlab-udc/lorehopqa", "lorepedia")     # background articles

item = qa["test"][0]
print(item["question"], "->", item["answer"])
for title, sentences in zip(item["context"]["title"], item["context"]["sentences"]):
    print(title, len(sentences))

To score with the official HotpotQA scorer, use the files under scorer/, which keep its field layout, rather than the data/ copies.

🛠️ Dataset Creation

Curation Rationale

An end-to-end answer score is evidence that better ranking improves generation only if the model cannot answer without the retrieved evidence. On Wikipedia-derived benchmarks that assumption frequently fails, so the score measures retrieval effectiveness and parametric fallback at once. Grounding the questions in invented worlds separates them.

Source Data

Items are built along two paths.

Structural transposition (7,371 items). A generator model receives only the structural schema of a HotpotQA item (passage counts, relevant passage indices, sentence lengths, supporting-fact positions and answer type), and writes a new item in an invented world. It never sees the original question, answer or text, so no content is paraphrased and only the shape of the reasoning transfers.
Instruction induction (13,241 items). Instructions induced from demonstrations generate items with evidence layouts the source collection does not contain: chains over three or four passages, convergent evidence, and candidate sets of six, eight or ten.

Quality Assurance

Candidates pass deterministic structural validation before any model judges them, then dual-juror verification: a quality juror scoring hop necessity, naturalness, distractor plausibility, answer precision and realism against a rubric, and a blind juror that receives the question with no context, any item it answers being discarded.

Construction is distributed across four open-weight LLM families with rotating generator/juror pairings, so no item is verified by the family that produced it.

🔒 Withheld Materials

The world specifications, the generation prompts and the per-item build record stay with the authors. The world text is what the documents were written from, and the build record links each transposed item to the HotpotQA item whose layout it reproduces; releasing either would undo the separation the collection depends on. Released identifiers are assigned independently of that record and the splits are shuffled, so neither the ids nor the file order recovers it. Attempting to reconstruct the withheld materials is prohibited by the agreement.

One further world, authored to the same specification, is withheld as a contamination probe.

⚠️ Bias, Risks, and Limitations

Synthetic content. Items are LLM-generated and LLM-verified. They carry the lexical and structural preferences of the four families that produced them, and acceptance was judged by jurors drawn from the same four.
Chance-answerable items. Boolean questions and comparisons that name every candidate are answerable above chance without context. Report metrics with these excluded as well as overall; the type and answer fields let you identify them.
Token-overlap metrics. Answer F1 measures overlap with a reference string, not correctness. Reference answers here are written as answers, so they admit less variation than spans of real text, which flatters F1 relative to HotpotQA.
Intra-world redundancy. An answer string can recur across documents of the same world, so a non-relevant same-world passage occasionally supports an answer without carrying the marked evidence. This is small but it is not zero.
Gating is not permanent protection. The property holds while evaluated models have not seen the collection. That is what the no-training term exists to preserve.

Recommendations

Report the chance-excluded condition alongside the overall one, and state the collection version used.

📈 First-Stage Retrieval

Rankings are shared across all readers. lore-* run over the test-split item documents, hotpot-* over the documents of the paired HotpotQA items. Adapted-embedder rows are reported in the paper.

System	R@2	R@10	all-gold@10
`hotpot-bm25`	0.556	0.785	0.591
`hotpot-dense-jina`	0.673	0.855	0.722
`hotpot-dense-minilm`	0.516	0.706	0.474
`hotpot-dense-simcse`	0.332	0.504	0.237
`hotpot-hybrid-jina`	0.677	0.858	0.726
`hotpot-hybrid-minilm`	0.546	0.770	0.571
`hotpot-hybrid-simcse`	0.406	0.660	0.413
`lore-bm25`	0.550	0.763	0.564
`lore-dense-jina`	0.501	0.707	0.459
`lore-dense-minilm`	0.421	0.617	0.345
`lore-dense-simcse`	0.230	0.389	0.157
`lore-hybrid-jina`	0.512	0.738	0.508
`lore-hybrid-minilm`	0.445	0.675	0.421
`lore-hybrid-simcse`	0.316	0.584	0.324

📝 Citation

See CITATION.cff in this repository. The paper is under review; the citation will be updated on publication.

👥 Dataset Card Authors

Juan Filgueiras, Marcos Fernández-Pichel and Javier Parapar (IRLab, CITIC, Universidade da Coruña).

📬 Contact

For questions and access requests: javier.parapar@udc.es

Downloads last month: 18

Total file size:

196 MB