---
pretty_name: HVMap Dataset
language:
- en
license: other
task_categories:
- text-retrieval
- text-ranking
tags:
- hvmap
- beir
- mteb
- text-retrieval
- text-ranking
- qrels
- federated-retrieval
- heterogeneous-vectordb
- vector-search
- cross-domain-retrieval
size_categories:
- 100K<n<1M
configs:
- config_name: all
  data_files:
  - split: corpus
    path:
    - data/corpus/fiqa.jsonl
    - data/corpus/bioasq.jsonl
    - data/corpus/scidocs.jsonl
    - data/corpus/scifact.jsonl
  - split: queries
    path: data/queries/queries.jsonl
  - split: qrels
    path: data/qrels/qrels.jsonl
- config_name: corpus
  data_files:
  - split: corpus
    path: data/corpus/*.jsonl
- config_name: finance
  data_files:
  - split: corpus
    path: data/corpus/fiqa.jsonl
- config_name: biomedical
  data_files:
  - split: corpus
    path: data/corpus/bioasq.jsonl
- config_name: science
  data_files:
  - split: corpus
    path:
    - data/corpus/scidocs.jsonl
    - data/corpus/scifact.jsonl
- config_name: queries
  data_files:
  - split: queries
    path: data/queries/queries.jsonl
- config_name: qrels
  data_files:
  - split: qrels
    path: data/qrels/qrels.jsonl
- config_name: fin_bio
  data_files:
  - split: corpus
    path:
    - data/corpus/fiqa.jsonl
    - data/corpus/bioasq.jsonl
  - split: queries
    path: data/queries/fin_bio.jsonl
  - split: qrels
    path: data/qrels/fin_bio.jsonl
- config_name: bio_sci
  data_files:
  - split: corpus
    path:
    - data/corpus/bioasq.jsonl
    - data/corpus/scidocs.jsonl
    - data/corpus/scifact.jsonl
  - split: queries
    path: data/queries/bio_sci.jsonl
  - split: qrels
    path: data/qrels/bio_sci.jsonl
- config_name: fin_sci
  data_files:
  - split: corpus
    path:
    - data/corpus/fiqa.jsonl
    - data/corpus/scidocs.jsonl
    - data/corpus/scifact.jsonl
  - split: queries
    path: data/queries/fin_sci.jsonl
  - split: qrels
    path: data/qrels/fin_sci.jsonl
- config_name: fin_bio_sci
  data_files:
  - split: corpus
    path:
    - data/corpus/fiqa.jsonl
    - data/corpus/bioasq.jsonl
    - data/corpus/scidocs.jsonl
    - data/corpus/scifact.jsonl
  - split: queries
    path: data/queries/fin_bio_sci.jsonl
  - split: qrels
    path: data/qrels/fin_bio_sci.jsonl
---

# HVMap Dataset

HVMap Dataset is a BEIR/MTEB-style text-retrieval benchmark package for
federated vector retrieval across heterogeneous vector databases. It contains
the retrieval corpus, mixed-domain queries, and qrels used by the HVMap
cross-domain benchmark described in *Federated Vector Retrieval across
Heterogeneous VectorDBs*.

## Repository Layout

| Path | Description |
| --- | --- |
| `data/corpus/*.jsonl` | Search corpus shards derived from FIQA, BioASQ, SciDocs, and SciFact. |
| `data/queries/*.jsonl` | Mixed-domain benchmark queries. `queries.jsonl` is the combined 2,000-query file. |
| `data/qrels/*.jsonl` | Qrels in JSONL form. `qrels.jsonl` is the combined qrels file. |
| `data/qrels/*.tsv` | BEIR/TREC-style qrels with header `query-id corpus-id score`. |
| `metadata/dataset_manifest.json` | Counts, source-corpus metadata, and validation summary. |

## Corpus

The corpus uses retrieval-standard fields:

| Field | Description |
| --- | --- |
| `_id` | Corpus document id. IDs are prefixed by domain, for example `fin-`, `bio-`, or `sci-`. |
| `title` | Source title, empty when not provided by the upstream corpus. |
| `text` | Searchable document text. |
| `source_dataset` | Local source dataset name: `fiqa`, `bioasq`, `scidocs`, or `scifact`. |
| `domain` | `finance`, `biomedical`, or `science`. |

Corpus size:

| Source dataset | Domain | Documents |
| --- | --- | ---: |
| `fiqa` | finance | 57,638 |
| `bioasq` | biomedical | 40,221 |
| `scidocs` | science | 25,657 |
| `scifact` | science | 5,183 |
| **Total** |  | **128,699** |

Upstream metadata fields that are not needed for retrieval are not included in
the corpus JSONL; the corpus is normalized to id/title/text plus source/domain
fields.

## Queries and Qrels

The benchmark contains 2,000 mixed-domain queries:

| Dataset alias | Qrels source | Queries | Qrels rows |
| --- | --- | ---: | ---: |
| `fin_bio` | `fin_bio_qrels.pickle` | 500 | 2,960 |
| `bio_sci` | `bio_sci_qrels.pickle` | 500 | 2,786 |
| `fin_sci` | `fin_sci_qrels.pickle` | 500 | 2,706 |
| `fin_bio_sci` | `fin_bio_sci_qrels.pickle` | 500 | 4,414 |
| **Total** |  | **2,000** | **12,866** |

The qrels files use binary score values. Each query-document link from the
source qrels is represented with `score = 1`.

Qrels JSONL fields:

| Field | Description |
| --- | --- |
| `qrel_id` | Stable row id in the form `{query_id}:{document_index}`. |
| `query_id` | Query id used in this package, such as `fin_bio:0000`. |
| `corpus_id` | Corpus document id matching the `_id` field in `data/corpus/*.jsonl`. |
| `score` | Binary qrels score. |
| `dataset_alias` | Mixed-domain dataset alias. |
| `qrels_name` | Source qrels pickle filename. |
| `query_index` | Zero-based query index within the source qrels file. |
| `document_index` | Zero-based linked-document position within that query. |

## Loading

Load the full benchmark:

```python
from datasets import load_dataset

dataset = load_dataset("snu-aidas/HVMap-dataset", "all")
corpus = dataset["corpus"]
queries = dataset["queries"]
qrels = dataset["qrels"]
```

Load only one mixed-domain benchmark subset:

```python
fin_bio = load_dataset("snu-aidas/HVMap-dataset", "fin_bio")
fin_bio_corpus = fin_bio["corpus"]
fin_bio_queries = fin_bio["queries"]
fin_bio_qrels = fin_bio["qrels"]
```

Available mixed-domain configs:

- `fin_bio`: finance + biomedical corpus, `fin_bio` queries/qrels
- `bio_sci`: biomedical + science corpus, `bio_sci` queries/qrels
- `fin_sci`: finance + science corpus, `fin_sci` queries/qrels
- `fin_bio_sci`: finance + biomedical + science corpus, `fin_bio_sci` queries/qrels

Load only a domain corpus:

```python
finance = load_dataset("snu-aidas/HVMap-dataset", "finance", split="corpus")
biomedical = load_dataset("snu-aidas/HVMap-dataset", "biomedical", split="corpus")
science = load_dataset("snu-aidas/HVMap-dataset", "science", split="corpus")
```

The `science` config reads both `scidocs` and `scifact` corpus shards. The
underlying files remain separate in the repository.

Load only combined queries or qrels:

```python
queries = load_dataset("snu-aidas/HVMap-dataset", "queries", split="queries")
qrels = load_dataset("snu-aidas/HVMap-dataset", "qrels", split="qrels")
```

For BEIR-style evaluation, download `data/corpus/*.jsonl`,
`data/queries/queries.jsonl`, and `data/qrels/qrels.tsv`. The TSV qrels files
use the conventional `query-id`, `corpus-id`, `score` header.

## Source Data and Licenses

This dataset contains mixed third-party source text. The repository metadata
therefore uses `license: other`; users must comply with the applicable upstream
licenses and attribution requirements.

| Upstream source | Domain role | License metadata |
| --- | --- | --- |
| [`BeIR/fiqa`](https://huggingface.co/datasets/BeIR/fiqa) | Financial | `cc-by-sa-4.0` |
| [`rag-datasets/rag-mini-bioasq`](https://huggingface.co/datasets/rag-datasets/rag-mini-bioasq) | Biomedical | `cc-by-2.5` |
| [`BeIR/scidocs`](https://huggingface.co/datasets/BeIR/scidocs) | Scientific | `cc-by-sa-4.0` |
| [`BeIR/scifact`](https://huggingface.co/datasets/BeIR/scifact) | Scientific | `cc-by-sa-4.0` |

The HVMap mixed-domain queries and qrels packaging are project-generated. The
embedded source corpus text remains subject to upstream dataset licenses.

## Validation

### Cross-Domain Benchmark Validation

Validation consists of a human relevance assessment and two independent
LLM-based consistency checks. For the human assessment, 100 queries were
randomly sampled from each of the four cross-domain settings, yielding 400
queries and 2,564 associated query-document pairs. Three authors independently
judged whether each document was relevant to its corresponding query.

| Validation source | # Queries | # Pairs | Relevant ratio |
| --- | ---: | ---: | ---: |
| Annotator 1 | 400 | 2,564 | 99.0% |
| Annotator 2 | 400 | 2,564 | 98.8% |
| Annotator 3 | 400 | 2,564 | 98.3% |
| *Mean pairwise agreement: 99.1%; Fleiss’ κ: 0.64* |  |  |  |
| Claude Opus 4.8 | 2,000 | 12,866 | 94.5% |
| Codex GPT-5.5 | 2,000 | 12,866 | 93.3% |

The human judgments yielded a mean pairwise agreement of 99.1% and a Fleiss’
κ of 0.64; the difference from the raw agreement reflects the highly imbalanced
label distribution. As a secondary full-benchmark consistency check, Claude
Opus 4.8 (max reasoning effort) and Codex GPT-5.5 (extra-high reasoning effort) independently judged all 12,866 query-document pairs across 2,000
queries. They judged 94.5% and 93.3% of the pairs as relevant, respectively.

### Corpus and Qrels Integrity

Before packaging, every qrels-linked `corpus_id` was checked against the
normalized corpus IDs included in this release.

- Corpus document IDs are unique: 128,699 unique IDs.
- All qrels-linked document IDs are present in the corpus.
- Missing qrels document count: 0.
- Qrels rows: 12,866.


## Intended Use

This dataset is intended for research on text retrieval, text ranking,
federated vector retrieval, heterogeneous vectorDB federation, cross-domain
retrieval, and benchmark construction.

## Limitations

- The qrels use binary scores.
- Some upstream corpus records have empty `text` fields in the local source
  corpus. These records are retained to preserve corpus id compatibility.
- The corpus is normalized for retrieval and omits heavy upstream metadata that
  is not required for retrieval experiments.

## Citation

If you use this dataset, cite the HVMap paper and the upstream corpora listed
above. A formal BibTeX entry can be added once the HVMap paper metadata is
finalized.