jeremycochoy/wikimedia-pageview-timeseries-raw

Wikimedia Pageview Time Series — full raw (wide format)

Full, unsampled Wikipedia pageview time series for every Wikimedia project (Wikipedia, Wiktionary, Commons, etc.), stored as raw wide parquet files: one row per article, one column per timestamp.

This is the complete derived output of the upstream pipeline — the companion repo jeremycochoy/wikimedia-pageview-timeseries holds a sampled, reshaped version (3.7 M rows in HF long format for training). Use this repo if you need the full data, or a resolution/decomposition the sampled repo doesn't expose.

The weekly stage-2 intermediate (fixed/, ~402 GB, 442 files) is not on HuggingFace — it is archived on private S3 Glacier at s3://frozen-datasets/wikimedia-pageview-hourly-2016W01-2024W25/. Everything derived from it is here.

Contents

Folder	Files	Size	Rows / file	Cols	Description
1024/	42 .parquet	~85 GB	~1.88 M	1028	Training-ready hourly windows of T=1025 consecutive hours (≈ 42.7 days). Covers 2011-12-01 → 2016-10-02.
1024-daily/	1 .parquet	3.2 GB	~1.99 M	1028	T=1025 daily windows (≈ 2.8 years per row).
1024-weekly/	1 .parquet	598 MB	~1.99 M	150	T=148 weekly windows (≈ 2.8 years per row), starting 2011-12-04.
1024-stl/<window>/	24 dirs × 3 = 72 .parquet	~551 GB	~1.88 M	1025	STL decomposition of the first 24 1024/ windows. Each dir has trend.parquet, seasonal.parquet, residual.parquet. Covers 2011-12-01 → 2014-08-19 (the last 18 windows were never decomposed).

Grand total: 116 files, ~640 GB.

Schema (wide format)

Rows are articles. Columns are:

1024/*.parquet and 1024-daily/*.parquet

Column	Type	Notes
index	int64	Numeric article id (stable across files)
key	string	Sanitised article id: <sanitised_domain_title>_<8char_sha256>
1025 datetime columns	float32	Hourly or daily view counts
__index_level_0__	int64	Pandas index artifact (can be dropped)

1024-weekly/*.parquet

Same as above but 148 datetime columns (weekly bucket) instead of 1025.

1024-stl/<window>/{trend,seasonal,residual}.parquet

Column	Type	Notes
index	int64
1024 datetime columns	float32	Hourly decomposed component. Note: no key column, and one fewer timestamp than 1024/ (STL trims the boundary).

How the data was produced

Three-stage pipeline in the open-source rnd repo:

1. scripts/pageview_loader.py — downloads hourly .bz2 files from dumps.wikimedia.org/other/pageview_complete/, filtering spider and automated traffic.
2. scripts/pageview_processor.py — decodes the Wikimedia letter- encoded hourly format (A–X = hours 0–23 followed by view count), filters pages with fewer than 10 daily views, writes one parquet per ISO week — produces the fixed/ intermediate (archived separately on private S3 Glacier, not in this repo).
3. scripts/pageview_weekly_to_fixed.py — stitches consecutive weeks from fixed/, cuts fixed-length windows — produces 1024/ (and, with modified T + resolution, 1024-daily/ / 1024-weekly/).

STL decomposition was produced by a separate Rust stage at scripts/stl_decomp/ in the same repo.

See docs/pageview-pipeline.md in the rnd repo for a full write-up.

Provenance timeline

Step	When
Raw .bz2 dumps downloaded	2024 (dumps later purged to save disk)
Weekly parquets (fixed/) produced	2024-05-20 → 2024-07-05
1024/, 1024-daily/, 1024-weekly/ produced	2024-06 → 2024-07
1024-stl/ decomposition (first 24 windows)	2024-06
Uploaded to this HF repo	2026-04

Loading

Raw wide parquets (this repo)

import pyarrow.parquet as pq
import numpy as np

table = pq.read_table("1024/2011-12-01.parquet")
df = table.to_pandas()

# 1025-long float32 windows, one row per article
values = df.iloc[:, 2:-1].values  # drop `index`, `key`, `__index_level_0__`
keys = df["key"].values
print(values.shape, values.dtype)  # (1877644, 1025) float32

Streaming via huggingface_hub

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="jeremycochoy/wikimedia-pageview-timeseries-raw",
    repo_type="dataset",
    local_dir="./wiki-raw",
    allow_patterns=["1024/*.parquet"],  # narrow to what you need
)

Reshape to HF long format

The companion repo jeremycochoy/wikimedia-pageview-timeseries stores the same data in long format (series: list[float32, 1025], source_id: uint8, meta: string). The transform lives in stage1_extract_wiki.py — _extract_series_matrix does the per-row wide-to-long conversion.

Known gaps

• STL coverage ends at 2014-08-19 — the last 18 1024/ windows (through 2016-10-02) were never decomposed. Regenerate via scripts/stl_decomp/ if needed.
• 1024/ windows cover 2011 → 2016 — they were produced from an earlier snapshot of fixed/ that no longer exists on disk. The current archived fixed/ only covers 2016-W01 → 2024-W25, so re-generating those specific 1024/ windows would require re-downloading pre-2016 raw dumps from Wikimedia.

License

Wikimedia pageview data is licensed CC0. This repo follows the upstream license.

Citation

If you use this data, please cite Wikimedia:

@misc{wikimedia_pageview_complete,
  author = {{Wikimedia Foundation}},
  title  = {Pageview complete dumps},
  year   = {2024},
  howpublished = {\url{https://dumps.wikimedia.org/other/pageview_complete/}},
}

Maintainer: jeremy@redstone.ee — issues / questions: open a discussion here or an issue in jeremycochoy/rnd.