README.md

---
configs:
- config_name: default
  data_files:
  - split: train
    path: "data/crawl=CC-MAIN-2026-25/**/*.parquet"
- config_name: CC-MAIN-2026-25
  data_files:
  - split: train
    path: "data/crawl=CC-MAIN-2026-25/**/*.parquet"
license: odc-by
task_categories:
- text-retrieval
language:
- multilingual
pretty_name: Common Crawl URL Seed
size_categories:
- 1B<n<10B
tags:
- common-crawl
- urls
- crawl-frontier
- parquet
- open-data
---

# **Common Crawl URL Seed**

> Every URL in a Common Crawl snapshot, as partitioned Parquet, ready to reseed a crawl

## What is it?

This dataset is the full URL list of a [Common Crawl](https://commoncrawl.org) snapshot, pulled straight from the crawl's columnar URL index and written as partitioned Parquet. Common Crawl is a non-profit that crawls the web and freely publishes its archives. Each snapshot ships a columnar index whose `url` column names every page the crawl captured. This dataset is that column, deduplicated and sharded, so a downstream crawler can reseed from a known frontier without pulling the 150 GB+ index again.

It currently holds crawl **CC-MAIN-2026-25** (warc subset) with **2,098,491,742 URLs across 256 shards**.

The dataset is released under the **Open Data Commons Attribution License (ODC-By) v1.0**, the same license Common Crawl uses.

## What is being released?

The URLs are sharded by host. Each shard is the set of URLs whose host hashes into one contiguous slice of the 64-bit hostkey space, so every URL of a host lands in exactly one shard and the shards tile the whole space with no gap and no overlap. That is the same partitioning [meguri](https://github.com/tamnd/meguri) assigns when it ingests a seed, so a shard maps one to one onto a crawl-frontier partition. The shards live under a crawl-partitioned directory:

```
data/
  crawl=CC-MAIN-2026-25/
    shard-00000.parquet
    shard-00001.parquet
    shard-00002.parquet
    ...
manifest.json
```

`manifest.json` is the shard map: it records each shard's hostkey range and row count, so a puller can rebuild the exact same frontier partitions the crawl used.

## Schema

| Column | Type | Description |
|---|---|---|
| `url` | string | A URL captured by the crawl |
| `host` | string | Host parsed from the URL, the key the shard is partitioned on |

## How to load it

### Using `datasets`

```python
from datasets import load_dataset

# stream every URL in the crawl
ds = load_dataset("open-index/commoncrawl-urls", name="CC-MAIN-2026-25", split="train", streaming=True)
for row in ds:
    print(row["url"])

# load one shard into memory
ds = load_dataset(
    "open-index/commoncrawl-urls",
    data_files="data/crawl=CC-MAIN-2026-25/shard-00000.parquet",
    split="train",
)
```

### Using DuckDB

```sql
SELECT url
FROM read_parquet('hf://datasets/open-index/commoncrawl-urls/data/crawl=CC-MAIN-2026-25/**/*.parquet')
WHERE host = 'en.wikipedia.org';
```

### Reseed a crawl frontier

Download the shards and hand them to meguri to rebuild the crawl frontier without touching Common Crawl again:

```bash
huggingface-cli download open-index/commoncrawl-urls --repo-type dataset \
  --include "data/crawl=CC-MAIN-2026-25/**" --local-dir ./cc-urls

# the parquet url column is the seed; feed it straight into a meguri store
meguri shard build --urls ./cc-urls/data/crawl=CC-MAIN-2026-25 --out ./frontier
```

## Provenance

The URLs come from the Common Crawl columnar URL index (`cc-index/table/cc-main/warc`), subset `warc`, for crawl `CC-MAIN-2026-25`. Only the `url` column of the index is read; nothing about page content is included here. The pull, the host sharding, and this dataset are produced by [ccrawl-cli](https://github.com/tamnd/ccrawl-cli).

## Sizes

Measured across 256 shards of CC-MAIN-2026-25.

| Stage | Size |
|---|---|
| URLs | 2,098,491,742 |
| URL text (uncompressed) | 149.4 GB |
| Parquet (Zstd) | 31.0 GB |

## Licensing

Released under the **Open Data Commons Attribution License (ODC-By) v1.0**. Use is also subject to [Common Crawl's Terms of Use](https://commoncrawl.org/terms-of-use). The URLs point at content that remains subject to its publishers' rights.