---
license: cc0-1.0
task_categories:
  - time-series-forecasting
tags:
  - time-series
  - wikimedia
  - pageviews
  - wikipedia
  - stl-decomposition
  - pretraining
size_categories:
  - 1M-10M
pretty_name: Wikimedia Pageview Time Series
dataset_info:
  - config_name: wiki_daily
    features:
      - name: series
        dtype: list
        list:
          dtype: float32
          length: 1025
      - name: source_id
        dtype: uint8
      - name: meta
        dtype: string
    splits:
      - name: train
        num_examples: 1990244
  - config_name: wiki_hourly
    features:
      - name: series
        dtype: list
        list:
          dtype: float32
          length: 1025
      - name: source_id
        dtype: uint8
      - name: meta
        dtype: string
    splits:
      - name: train
        num_examples: 3715121
  - config_name: wiki_stl_residual
    features:
      - name: series
        dtype: list
        list:
          dtype: float32
          length: 1025
      - name: source_id
        dtype: uint8
      - name: meta
        dtype: string
    splits:
      - name: train
        num_examples: 530731
  - config_name: wiki_stl_seasonal
    features:
      - name: series
        dtype: list
        list:
          dtype: float32
          length: 1025
      - name: source_id
        dtype: uint8
      - name: meta
        dtype: string
    splits:
      - name: train
        num_examples: 371512
  - config_name: wiki_stl_trend
    features:
      - name: series
        dtype: list
        list:
          dtype: float32
          length: 1025
      - name: source_id
        dtype: uint8
      - name: meta
        dtype: string
    splits:
      - name: train
        num_examples: 159219
configs:
  - config_name: wiki_daily
    data_files: "wiki_daily/*.parquet"
  - config_name: wiki_hourly
    data_files: "wiki_hourly/*.parquet"
  - config_name: wiki_stl_residual
    data_files: "wiki_stl_residual/*.parquet"
  - config_name: wiki_stl_seasonal
    data_files: "wiki_stl_seasonal/*.parquet"
  - config_name: wiki_stl_trend
    data_files: "wiki_stl_trend/*.parquet"
---

# Wikimedia Pageview Time Series

Preprocessed time series dataset derived from [Wikimedia pageview statistics](https://dumps.wikimedia.org/other/pageview_complete/). Contains fixed-length windows of Wikipedia article pageview counts at hourly and daily resolution, plus STL seasonal-trend decomposition components.

## Dataset Summary

| Subset | Series Count | Series Length | Size | Description |
|---|---|---|---|---|
| `wiki_hourly` | 3,715,121 | 1025 | 2.5 GB | Hourly pageview counts |
| `wiki_daily` | 1,990,244 | 1025 | 2.5 GB | Daily aggregated pageview counts |
| `wiki_stl_residual` | 530,731 | 1025 | 2.0 GB | STL decomposition — residual component |
| `wiki_stl_seasonal` | 371,512 | 1025 | 1.4 GB | STL decomposition — seasonal component |
| `wiki_stl_trend` | 159,219 | 1025 | 587 MB | STL decomposition — trend component |

**Total: ~6.77M time series, ~9 GB**

## Schema

Each parquet file contains three columns:

| Column | Type | Description |
|---|---|---|
| `series` | `fixed_size_list<float32>[1025]` | The time series values (1024 input steps + 1 target) |
| `source_id` | `uint8` | Numeric identifier for the data source/component |
| `meta` | `string` | Human-readable component name |

### Source IDs

| `source_id` | `meta` value | Description |
|---|---|---|
| 1 | `wiki_hourly` | Raw hourly pageview counts |
| 2 | `wiki_daily` | Daily aggregated pageview counts |
| 3 | `wiki_stl_residual` | Residual after STL decomposition |
| 4 | `wiki_stl_seasonal` | Seasonal component from STL decomposition |
| 5 | `wiki_stl_trend` | Trend component from STL decomposition |

## Data Origin

- **Source:** [Wikimedia pageview complete dumps](https://dumps.wikimedia.org/other/pageview_complete/)
- **Date range:** December 2011 — October 2016
- **Filtering:** Pages with fewer than 10 daily views are excluded
- **Processing pipeline:**
  1. Raw hourly `.bz2` dumps downloaded from Wikimedia
  2. Parsed and aggregated into weekly parquet files
  3. Stitched into fixed-length windows of T=1025 time steps (1024 + 1)
  4. STL seasonal-trend decomposition applied to extract trend, seasonal, and residual components
  5. Daily aggregation computed from hourly data

## Usage

```python
from datasets import load_dataset

# Load a specific subset
ds = load_dataset("jeremycochoy/wikimedia-pageview-timeseries", "wiki_daily")

# Access a time series
series = ds["train"][0]["series"]  # list of 1025 floats
```

Or load directly with PyArrow:

```python
import pyarrow.parquet as pq

table = pq.read_table("wiki_daily/wiki_daily_file000_00000.parquet")
df = table.to_pandas()
series = df["series"].iloc[0]  # numpy array of shape (1025,)
```

## Data Characteristics

- **Patterns present:** viral spikes, seasonal cycles (holidays, sports events, school calendars), slow decays, flat/stable pages, multi-language diversity (all Wikimedia projects)
- **Languages:** All Wikimedia language editions included (English, German, French, Japanese, Russian, etc.)
- **Use cases:** Time series foundation model pretraining, forecasting benchmarks, transfer learning

## License

The underlying Wikimedia pageview data is released under [CC0 1.0 (Public Domain)](https://creativecommons.org/publicdomain/zero/1.0/). This preprocessed dataset inherits that license.