Wanderlens

The data behind Wanderlens, a trip search that finds destinations by what a traveler describes ("warm sea, empty beaches, and nothing to do but read") rather than by place name.

It holds 10,973 Wikivoyage destinations cut into 138,753 retrieval chunks, labels for each place, credits for an openly licensed photo of most places, the chunk embeddings the site searches, and the synthetic queries used to fine-tune armkh/wanderlens-embed.

Files

Path	Rows	What it is
places/destinations.parquet	10,973	One row per destination: page_id, title, Wikivoyage url, wikidata_id, type (city, region, park, rural area), guide quality, lat/lon, country, region_path, abstract, date_modified
places/sections.parquet	118,563	Guide sections as clean text (Understand, See, Do, Eat, Drink, Sleep…), with heading and position
places/district_sections.parquet	6,291	Sections of district pages, attached to their parent city
places/chunks.parquet	138,753	Retrieval chunks: chunk_id ({page_id}-{n}), section, heading, context (place and section), text
places/enrichment.parquet	10,973	Labels per place: best_months, avoid_months, months_source, budget, vibes, settings, tagline, model
places/images.parquet	8,737	One lead photo per place: Wikimedia image_url, Commons image_page, author, license, license_url
embeddings/chunks_bge-small-ft.parquet	138,753	chunk_id and a 384-dim normalized vector from the fine-tuned model
embeddings/chunks_bge-small.parquet	138,753	The same chunks embedded with the base BAAI/bge-small-en-v1.5, for comparison
training/queries.parquet	32,917	Three synthetic queries per destination (specific, casual, multi) with the positive chunk each was written from
training/splits.parquet	10,973	train (9,473), dev (500) or test (1,000) per destination
training/triplets_bge-small.parquet	28,401	Training triplets: query, positive chunk, and a hard negative mined with the base model
eval/metrics.json		Retrieval metrics for the base and fine-tuned models (bge-small and bge-base)
eval/metrics_onnx.json		The same metrics for the int8 ONNX exports the site runs
eval/hand_written_queries.json		32 hand-written queries with the text patterns a relevant guide must contain

A chunk is embedded as context + "\n" + text. Queries take the bge instruction prefix Represent this sentence for searching relevant passages: .

How it was built

1. Source. Wikivoyage pages from Wikimedia's linked OpenStreetMap / Wikidata / Wikipedia / Wikivoyage dataset, parsed with mwparserfromhell into sections and chunks.
2. Labels. gpt-5.4-nano (OpenAI Batch API) read each guide and returned travel months, a budget level, vibes, settings and a one-line tagline. Most month labels (9,156) come from the model's general knowledge because the guide doesn't say; months_source records which.
3. Photos. Each place's lead image from the Wikimedia API, kept only when it is a photo (not a map) under CC BY, CC BY-SA, CC0 or public domain, with an author where the licence requires one.
4. Queries. gpt-5.4-nano wrote three traveler-style queries per destination from its guide text. 22 queries name the place itself (leaked); they are kept but flagged.
5. Splits. Destinations, not queries, are split, so test places and their text never appear in training.

The code for every step is in the GitHub repository.

Licence and attribution

Travel text is from Wikivoyage contributors and is licensed CC BY-SA 4.0; this dataset, which is derived from it, uses the same licence. Each destination's url links to the page and its history of authors.

places/images.parquet contains links and credits, not images. Each photo is hosted on Wikimedia Commons under the licence given in its row, and must be credited to its author when used.

The labels and queries were generated by an OpenAI model and may contain mistakes.