Arctic-Wiki-English-5M

VDBBench-compatible vector benchmark case published as a Hugging Face dataset repository.

What’s in this repo

• vdbbench/: a Parquet-only folder intended to be downloaded as-is for VDBBench.
• The Dataset Viewer indexes only train/test (and optionally shuffled), and intentionally ignores neighbors.parquet.

vdbbench/ file contract (VDBBench)

• Train vectors: train.parquet or train-*-of-*.parquet
• Query vectors: test.parquet or test-*-of-*.parquet
• Ground truth: neighbors.parquet with columns id, neighbors_id (list[int])
• Optional shuffled train: shuffle_train.parquet or shuffle_train-*-of-*.parquet

Important: do not add non-parquet files inside vdbbench/ (VDBBench requires the folder to be parquet-only).

Download for VDBBench

Download only the parquet files and point VDBBench at the vdbbench/ folder.

hf download "AI71ai/Arctic-Wiki-English-5M" \
  --repo-type dataset \
  --include "vdbbench/*.parquet" \
  --local-dir ./my_case

Then set VDBBench /custom → Folder Path to:

• ./my_case/vdbbench

How this case was created

• Source: wikimedia/wikipedia at revision b04c8d1ceb2f5cd4588862100d08de323dccfbaa (pinned for reproducibility when set).
• Embedding model: Snowflake/snowflake-arctic-embed-l-v2.0 (dimension 1024).
• Languages: en
• Train size: 5,000,000 vectors
• Test size: 1,000 query vectors
• Phase 2 seed: 42

Size filtering (chosen to embed well)

We filter Wikipedia articles by character length of the article text to avoid pages that are too short (low-signal) or too long (risking truncation / poor embedding behavior).

Approx token ranges below are computed from the Snowflake Arctic tokenizer’s observed chars/token ratios (using the per-language ~8k-token max lengths we chose).

Language	Text length filter (chars)	Approx token range
en	400–32,605 chars	~98–~8,000 tokens

Test set (queries)

• Queries are derived from Wikipedia titles of documents that appear in the training stream.
• Phase 2 Stage 1 selects test_size titles using reservoir sampling (bounded memory) while writing train.
• Phase 2 Stage 2 embeds those titles to produce test*.parquet (query vectors).

Shuffled train (optional)

• Some vector DBs are sensitive to ingestion order; this benchmark optionally provides a shuffled training set to evaluate that behavior.
• This repo includes shuffle_train*.parquet for VDBBench “Use Shuffled Data” mode.

Ground truth (neighbors.parquet)

• neighbors.parquet contains the exact top-400 nearest train IDs for each test query ID.
• Similarity is computed with cosine (implemented as L2-normalize + inner product).

Provenance: ID map sidecars (maps/)

This repo also includes optional ID mapping sidecars under maps/ that let you trace VDBBench IDs back to the original Wikipedia/source IDs used during embedding.

• maps/ is intentionally not part of any Dataset Viewer split.
• Files are Parquet (same sharding convention as vectors when applicable).

Files

• train_id_map*.parquet: train VDBBench id → source id mapping
• test_id_map*.parquet: test query id → train_id + source id mapping
• shuffle_train_id_map*.parquet: shuffled-train VDBBench id → source id mapping

Schemas

• train_id_map: id, source_id
• test_id_map: id, train_id, source_id
• shuffle_train_id_map: id, source_id

Notes:

• source_id refers to the upstream article ID from the embedded Wikipedia dump (see the pipeline repo for details).
• For mixed-language cases, the language column indicates which language shard the source came from.

Design decisions

• Per-case repos: keeps each benchmark case independently versionable and easily downloadable without filtering huge multi-case repos.
• Parquet-only vdbbench/ folder: matches VDBBench’s strict folder selection rule and enables “download folder → run benchmark”.
• Neighbors kept next to vectors, but not indexed: neighbors.parquet is required by VDBBench, but it’s not useful to display as a dataset split in the viewer.
• Prefer single-file test.parquet: VDBBench duplicates test queries across processes; keeping queries small and single-file reduces overhead.
• Embedding model choice: Snowflake Arctic Embed L v2.0 is high on the MTEB-R leaderboard (see https://huggingface.co/spaces/mteb/leaderboard) and is a strong fit for evaluating compressed representations (including binarized/quantized vectors).

Usage with datasets

from datasets import load_dataset

ds = load_dataset("AI71ai/Arctic-Wiki-English-5M")
print(ds)

Note: neighbors.parquet is intentionally not part of any split.