QUASR Database: Diverse Research Repositories

Updated 5 July 2026

QUASR Database is a set of independently developed, domain-qualified repositories spanning question answering, speech processing, and stellarator design with unique data structures.
The heterogeneous QA resource employs multi-modal retrieval from text, tables, and knowledge graphs, achieving state-of-the-art precision improvements in answer generation.
Additional QUASR-labeled databases include QA-by-search datasets, an annotated Arabic speech corpus, and a stellarator design archive, each optimized for its specific research application.

“QUASR Database” is not a single, universally standardized object in the research literature. In arXiv usage, the label and closely related spellings refer to several distinct database-like resources: a retrieval-augmented question-answering system over heterogeneous sources; a pair of “Question Answering by Search and Reading” datasets; a large-scale Arabic speech corpus; and the QUAsi-symmetric Stellarator Repository. These resources share a database orientation—indexed evidence, reusable corpora, or machine-readable repositories—but they belong to different technical domains and should not be conflated (Christmann et al., 2024, Dhingra et al., 2017, Mubarak et al., 2021, Giuliani et al., 2024).

1. Nomenclature and referential scope

In current usage, “QUASR Database” most often requires disambiguation by domain. The following senses are explicitly attested in the cited literature.

Name	Domain and contents	arXiv id
QUASAR / QUASR	RAG-based question answering over unstructured text, structured tables, and knowledge graphs	(Christmann et al., 2024)
Quasar	QA-by-search-and-reading datasets: Quasar-S and Quasar-T	(Dhingra et al., 2017)
QASR	Multi-dialect Arabic broadcast speech corpus with aligned transcriptions and metadata	(Mubarak et al., 2021)
QUASR	QUAsi-symmetric Stellarator Repository with stellarator configurations and coil sets	(Giuliani et al., 2024)
QUADRo	Related q/a-pair retrieval database for open-domain QA	(Campese et al., 2023)

The principal source of ambiguity is orthographic rather than conceptual. “QUASAR,” “QUASR,” “Quasar,” and “QASR” denote unrelated systems in question answering, speech technology, and plasma physics. A common misconception is that the term names a single cross-domain database. The literature instead shows several independent resources with different objectives, data models, and evaluation protocols.

This multiplicity also changes the meaning of “database.” In the heterogeneous QA system, the term denotes a virtual database assembled from text, tables, and knowledge graphs. In the 2017 Quasar work, it denotes a research environment for retrieval plus reading. In QASR, it denotes an annotated corpus and metadata resource. In the stellarator repository, it denotes a large scientific design archive. A plausible implication is that “database” here is best understood functionally—as an indexed, queryable research substrate—rather than as a single relational schema (Campese et al., 2023).

2. QUASAR/QUASR as a heterogeneous question-answering database

In "RAG-based Question Answering over Heterogeneous Data and Text" (Christmann et al., 2024), QUASAR, also written QUASR in some contexts, is a retrieval-augmented question-answering system over unstructured text, structured tables, and knowledge graphs. Its stated goal is to act like a database over heterogeneous information. The system takes a natural-language question, derives a Structured Intent (SI), retrieves evidence from Wikidata, all English Wikipedia articles, and all Wikipedia tables including infoboxes, re-ranks and filters the evidence, and then generates an answer with Llama-3.1-8B-Instruct.

The architecture has four sequential stages: Question Understanding (QU), Evidence Retrieval (ER), Re-Ranking and Filtering (RF), and Answer Generation (AG). QU maps the question to SI, a faceted frame with slots such as Ans-Type, Entities, Relation, Time, and Location. The QU module uses BART-base and reuses 3,400 silver-standard (question, SI) pairs from the CompMix benchmark. ER then collects up to about 1000 candidate evidence pieces from the knowledge graph, text, and tables. RF reduces this pool to about 30 items by either GNN-based RF or CE-based RF. AG consumes SI plus the top evidence and produces a grounded natural-language answer.

A defining feature is the unified treatment of heterogeneous data. Knowledge-graph triples with qualifiers, table rows with column headers and DOM context, and text sentences are all converted into textual pseudo-sentences. BM25 is used for initial scoring of text and table rows, while KG retrieval uses CLOCQ to obtain a relevant entity-centric subgraph. After verbalization, all evidence is pooled and ranked globally. This design permits source-specific retrieval while maintaining a common downstream representation.

The paper reports that heterogeneous fusion is materially important. On CompMix, P@1 is 0.455 for text only, 0.481 for KG only, 0.432 for tables only, and 0.564 for text+KG+tables. Re-ranking is likewise central: on CompMix dev, using only BM25 top-30 gives P@1 = 0.468, whereas GNN 1000→100→30 gives 0.574 and CE 1000→100→30 gives 0.573. On the main benchmark table, QUASAR reaches 0.564 on CompMix and 0.754 on TimeQuestions, with the latter reported as a new state of the art. It uses about 8,218M parameters in total, compared with 175,000M for GPT-3 and an estimated 1,760,000M for GPT-4, and the paper estimates computational cost reductions of about 200× versus GPT-3 and 2000× versus GPT-4 under a simple “cost ∝ #parameters” assumption.

From a data-management perspective, the system resembles a logical mediator over multiple physical sources. Retrieval over BM25 indexes and CLOCQ is separated from answer generation, evidence can be surfaced as explanations, and periodic re-indexing of Wikipedia and Wikidata would refresh the underlying knowledge substrate. The paper does not define an explicit external API, but its internal decomposition is sufficiently structured that one could expose each stage independently. This suggests a database interpretation in which natural-language querying is compiled into SI-guided access over heterogeneous indexed stores rather than executed against a single monolithic model.

3. Quasar as a QA-by-search-and-reading research database

The earlier "Quasar: Datasets for Question Answering by Search and Reading" (Dhingra et al., 2017) defines Quasar not as a single inference system but as a pair of large-scale QA datasets: Quasar-S and Quasar-T. The resource is explicitly designed for end-to-end study of two coupled subtasks: search for relevant text in a large corpus, and read that text to extract the answer. It therefore functions as a database-like experimental environment for retrieval plus machine reading.

Quasar-S is a closed-domain cloze dataset about software, built from Stack Overflow tag excerpts. It contains 37,362 cloze questions after filtering, with splits of 31,049 train, 3,174 validation, and 3,139 test. All answers come from a fixed vocabulary of 4,874 Stack Overflow tags. The background corpus is constructed from Stack Overflow by scraping the top 50 question threads for each entity, removing code, segmenting sentences, and building short and long pseudodocuments.

Quasar-T is an open-domain trivia dataset grounded in ClueWeb09. The final splits are 37,012 train, 3,000 validation, and 3,000 test. Because answers are free text, each question is accompanied by a question-specific candidate set derived from noun phrases in the retrieved context, with the gold answer added if necessary. ClueWeb09 pages are parsed into short and long pseudodocuments, again yielding a search-and-read setup rather than a fixed-passage reading-comprehension setup.

The database aspect is operationalized through explicit retrieval infrastructure. Quasar-S indexes Stack Overflow pseudodocuments with Apache Lucene and uses a query combining phrase match, bag-of-words match, and a mandatory head-tag constraint. Quasar-T first retrieves 100 HTML pages per question from ClueWeb09 with Indri, then builds an in-memory Lucene index over pseudodocuments for second-stage ranking. The release includes corpora, retrieval code, pre-retrieved contexts, and candidate lists.

Evaluation separates Search Accuracy, Reading Accuracy, and overall performance. For Quasar-S, the best automatic system lags human performance by 16.4 percentage points. For Quasar-T, the best reported automatic result is 0.285 F1, which is 32.1 percentage points below human F1 of about 0.606. The paper emphasizes a trade-off already familiar in later RAG work: increasing the number of retrieved documents improves answer coverage but can reduce reader effectiveness as irrelevant context grows.

Historically, this resource is important because it prefigures later retrieval-augmented QA formulations while preserving explicit control over retrieval. Unlike systems tied to commercial search engines, Quasar exposes negative contexts, retrieval outputs, and standardized corpora. In that sense, it established a database-style research protocol for QA before contemporary large-language-model RAG became standard.

4. QASR as an Arabic speech and speech-NLP database

In speech technology, "QASR: QCRI Aljazeera Speech Resource -- A Large Scale Annotated Arabic Speech Corpus" (Mubarak et al., 2021) is a large multi-dialect Arabic broadcast corpus that is sometimes informally called a “QUASR database.” It is built from the Aljazeera Arabic news archive and contains 2,041 hours of speech, 3,545 episodes, and about 1.6 million segments, sampled at 16 kHz. The corpus also includes a 130M-word text dataset for language modeling.

QASR is organized around lightly supervised transcriptions aligned to audio by a modified Smith–Waterman procedure. It stores segment-level start and end times, speaker IDs, timed word sequences, and quality indicators including Average Word Duration (AWD), Grapheme Match Error Rate (GWER), and Word Match Error Rate (WMER). The segmentation is explicitly linguistically and acoustically motivated: punctuation, speaker turns, short pauses, POS tags, NER tags, and a manually pruned NO-STOP-LIST are used to avoid cutting inside named entities and noun phrases.

The annotation layer goes beyond conventional ASR corpora. QASR provides punctuation, segment-level speaker information, standardized speaker names, gender for about 2,000 common speakers covering approximately 82% of all segments, country labels for test speakers, code-switching tags, and NE annotations on the shared MGB-2 test set. The corpus is therefore positioned as a speech-plus-NLP resource rather than only an ASR training set.

The paper reports baseline results for several tasks. An end-to-end Transformer trained on QASR achieves 15.1% dev WER and 14.7% test WER on MGB-2, close to a carefully tuned model trained on MGB-2 itself. For punctuation restoration, a Transformer + BiLSTM model using AraBERT obtains on the test set F1(O) = 98.3%, F1(COMMA) = 46.7%, F1(FSTOP) = 59.7%, and F1(QUES) = 54.9%. The corpus is also used for speaker verification, dialect identification, NER on speech transcripts, and code-switching analysis.

As a database, QASR is defined by the joint availability of audio, aligned transcripts, metadata, and downstream-task annotations. Its practical significance lies in making broadcast Arabic queryable at multiple representational levels: waveform, segment, word, speaker, punctuation, dialect, and named entity. This suggests a markedly different notion of “database” from QA resources: here the central object is a richly indexed multimodal corpus for supervised and weakly supervised spoken-language research.

5. QUASR as the QUAsi-symmetric Stellarator Repository

In plasma physics, QUASR denotes the QUAsi-symmetric Stellarator Repository, a large database of vacuum-field stellarator configurations with associated filamentary coil sets (Giuliani et al., 2024). The 2024 augmentation expands the repository to include quasihelically symmetric stellarators and reports a database of almost 370,000 quasiaxisymmetric and quasihelically symmetric devices. The repository is archived at 10.5281/zenodo.10050655 and can be explored online at quasr.flatironinstitute.org.

Each device is defined by a particular coil set with currents producing a three-dimensional magnetic field and nested flux surfaces over some radial extent. Stored quantities include Boozer-compatible surface parameterizations, magnetic-axis geometry, rotational transform profiles and mean values, aspect ratio, the number of field periods, quasisymmetry metrics, and coil-quality metrics. The database enforces or records constraints such as $\kappa_{\max} \le 5\,\text{m}^{-1}$ and $\langle \kappa^2 \rangle \le 5\,\text{m}^{-2}$ , together with coil–plasma distance, coil–coil spacing, and total coil length targets.

The repository is produced by a globalized multi-phase optimization workflow. A TuRBO globalization layer samples coil parameter space. Phase I uses near-axis expansion coil design, Phase II applies BoozerLS, and Phase III applies BoozerExact. For interpretation, the paper develops two exploration methods: near-axis quasisymmetry landscape mapping and principal component analysis over 663-dimensional feature vectors of Fourier surface coefficients. For many subsets, one to three principal components are sufficient to characterize most of the variance.

The structure revealed is not random. QA devices cluster in a central QA region of the near-axis landscape, while QH devices occupy distinct QH phases associated with different helicities. PCA shows branches, clusters, and low-dimensional manifolds in fixed- $(n_{\text{fp}}, \bar\iota, \text{QS type})$ subsets. Previously published devices such as HSX, WISTELL-A, PRECISE_QH, PRECISE_QA, BULLER, NIES configurations, and GIU_QA can be projected into the same low-dimensional spaces and compared directly.

In this context, “database” refers to a systematic scientific design archive rather than an information-retrieval backend. Its function is to support stellarator optimization, theory benchmarking, and data-driven modeling by exposing a large, internally consistent collection of geometry, coil, and quasisymmetry data. The repository thereby turns configuration space itself into a searchable object.

A closely related but distinct resource is QUADRo, introduced in "QUADRo: Dataset and Models for QUestion-Answer Database Retrieval" (Campese et al., 2023). QUADRo is not named QUASR, but it is directly relevant because it operationalizes a question–answer database as the primary knowledge source for open-domain QA. The database contains about 6.3M English QA pairs and uses a dense neural retriever plus a cross-encoder reranker. The retrieval model encodes a target question against indexed concatenations of stored question–answer pairs, and the best reranking configuration, Electra $_{QAQ}$ , reaches P@1 = 50.8 on the ranking dataset. In end-to-end comparisons, QUADRo reports 58.0 on Quora, 50.5 on NQ, and 29.3 on TriviaQA, versus 35.0, 56.0, and 27.0 for a WebQA baseline.

This comparison helps clarify several recurring confusions. First, QUASAR/QUASR in the heterogeneous RAG paper is not a q/a-pair retrieval database in the QUADRo sense: it retrieves evidence snippets from text, tables, and KGs, not previously answered questions. Second, the 2017 Quasar datasets are benchmarks and retrieval environments, not an answer-serving system. Third, QASR the Arabic speech corpus and QUASR the stellarator repository are unrelated to question answering. The shared acronym therefore does not imply shared architecture, shared data model, or shared evaluation methodology.

A further misconception is that “database” implies a stable schema and identical access pattern across these resources. The literature shows the opposite. In QA, the dominant access pattern is natural-language querying over indexed evidence or indexed q/a pairs. In speech, it is corpus access over aligned multimodal annotations. In stellarator design, it is filtering and projecting high-dimensional physical configurations. This suggests that the most precise use of “QUASR Database” is always domain-qualified: QUASAR/QUASR for heterogeneous QA, Quasar for QA-by-search-and-reading datasets, QASR for Arabic speech, and QUASR for the stellarator repository.