README.md

---
license: cc-by-sa-4.0
language:
  - en
pretty_name: English Wikivoyage Retrieval 2026-05
task_categories:
  - sentence-similarity
  - feature-extraction
  - question-answering
tags:
  - wikivoyage
  - wikimedia
  - retrieval
  - vector-search
  - lancedb
  - faiss
  - sqlite
  - travel
size_categories:
  - 100K<n<1M
configs:
  - config_name: corpus_chunks
    default: true
    data_files:
      - split: train
        path: parquet/chunks/*.parquet
  - config_name: travel_scenarios
    data_files:
      - split: test
        path: data/scenarios/travel_scenarios_public.parquet
  - config_name: retrieval_eval
    data_files:
      - split: test
        path: eval/results/retrieval_eval_public.parquet
---

# English Wikivoyage Retrieval 2026-05

I like travel datasets because they are about real places, real constraints,
and the small practical questions people ask before they go somewhere. I am
sharing this English Wikivoyage retrieval corpus in that spirit: as honest
work from a researcher-builder who wants to explore the world, make the
pipeline inspectable, and let other people reuse or challenge the choices.

This is not presented as a finished travel product or a private leaderboard.
It is a concrete public artifact: text chunks, provenance, retrieval pools,
scenario probes, and serving files that can be opened, measured, and rebuilt.

The dataset is built for travel-oriented retrieval experiments. It includes a
cleaned and chunked Wikivoyage text corpus, explicit retrieval pools, public
scenario probes, and serving artifacts for dense, lexical, and hybrid search.
The first companion use case is a Hugging Face Space for destination and travel
search demos:

```text
vincentss/wikivoyage-playground
```

## What To Look At First

Start with the `corpus_chunks` viewer table. It is the default Dataset Viewer
config and shows the actual text chunks, page and section provenance,
retrieval pool assignment, source URL, and public quality flags. This is the
canonical corpus artifact. The Lance, SQLite, and FAISS files are serving
artifacts derived from this corpus.

Then look at `travel_scenarios` and `retrieval_eval` to see the question shapes
used to probe the system and how the released retrieval routes behaved.

## Research Angle

The main contribution is not simply embedding Wikivoyage. The useful part is
the travel-aware data shape: page cleaning, section-aware chunks,
listing-aware metadata, retrieval pools, and scenario probes that reflect how
people ask travel questions.

The release is intentionally experimental. It is meant to be realistic enough
to run a demo or rebuild a retriever, while staying honest about limits:
English Wikivoyage only, lightweight public probes, no media embeddings, and no
claim that it replaces current official travel guidance.

## Contents

The v2 release is organized around three inspectable dataset configs:

| Config | File(s) | Purpose |
| --- | --- | --- |
| `corpus_chunks` | `parquet/chunks/*.parquet` | Canonical public chunk corpus for inspection and re-embedding. |
| `travel_scenarios` | `data/scenarios/travel_scenarios_public.parquet` | Realistic public travel queries and intended route pools. |
| `retrieval_eval` | `eval/results/retrieval_eval_public.parquet` | Public aggregate retrieval metrics by backend/route. |

It also includes practical serving artifacts:

- `lance/bge_m3_1024.lance`
- `indexes/sqlite_fts5/`
- `indexes/faiss/bge_m3_1024/`
- `data/scenarios/travel_scenarios_public.jsonl`
- `eval/results/retrieval_eval_public.jsonl`
- `data/schema/chunks_schema.json`
- `data/schema/retrieval_eval_schema.json`
- `space/contract.json`
- `manifest.json`
- `checksums.sha256`
- `LICENSE_DATA`

See `manifest.json` for exact file sizes, checksums, row counts, and release
lineage.

## Source And Attribution

Source project: English Wikivoyage.

Source data comes from public Wikimedia project dumps. Row-level artifacts
preserve page title, source URL, and revision URL where available. This dataset
is not affiliated with or endorsed by the Wikimedia Foundation.

Source text is copyright Wikimedia contributors and is available under the
Creative Commons Attribution-ShareAlike 4.0 International License unless
otherwise noted. Dataset artifacts derived from Wikivoyage text, including
embeddings and search indexes, are released under CC BY-SA 4.0.

See `LICENSE_DATA` for links and attribution guidance.

## Intended Use

This dataset is intended for:

- travel search demos;
- vector, lexical, and hybrid retrieval experiments;
- LanceDB and local-index provisioning examples;
- retrieval-augmented answer prototypes over Wikivoyage content;
- research inspection of travel-aware chunking and retrieval routing.

It is not intended for:

- safety-critical travel advice;
- replacing current official travel guidance;
- broad Wikimedia corpus claims;
- media/image retrieval unless a future release explicitly includes media
  artifacts.

## Known Limitations

- English Wikivoyage only.
- Scenario checks are realistic probes, not exhaustive relevance judgments.
- Listing chunks intentionally dominate raw row count, so retrieval should route
  by `retrieval_pool` and apply caps/diversity for broad destination queries.
- Media/image embeddings are not included.
- Remote query performance depends on Hugging Face Hub and Space hardware
  limits.

## Citation

If this dataset is useful, please cite the dataset repository and attribute the
underlying Wikivoyage source pages through the row-level URLs. If you
redistribute modified versions, preserve compatible CC BY-SA attribution and
share-alike terms.