davanstrien/commoncrawl-jobs-demo

Common Crawl on Jobs — datatrove JobsPipelineExecutor demo

228,088 English web documents (~820 MB compressed, 60 jsonl.gz shards) extracted from 1,237,374 Common Crawl pages — the output of a test run of datatrove's experimental JobsPipelineExecutor, which fans a datatrove pipeline out across a pool of Hugging Face Jobs instead of a Slurm cluster.

This is a pipeline demo artifact, not a curated corpus: one segment slice of one crawl, shared as the verifiable receipt for the run's numbers. For a real FineWeb-style corpus, use HuggingFaceFW/fineweb.

How it was made

Source: WARCs 00000–00059 of crawl CC-MAIN-2026-25, segment 1780687572080.85, streamed directly from Common Crawl's Hugging Face bucket — no download step.

Pipeline (the standard FineWeb extraction shape, one WARC per Job, 60 × cpu-basic):

WarcReader → URLFilter → Trafilatura(timeout=10) → LanguageFilter → GopherRepetitionFilter → GopherQualityFilter → JsonlWriter

stage	documents
pages read	1,237,374
reached Gopher quality filter	822,310
written	228,088

Total compute: 93.87 Job-hours on cpu-basic ($0.01/hour) = $0.94 — including every failed attempt during testing; the 60 ranks that produced the data used 48.79 hours = $0.49. Exact per-job records in the cost receipts below. Per-rank wall time roughly doubles at 60-way fan-out versus low concurrency (shared-CPU co-tenancy), and since billing is per job-hour that roughly doubles cost per page too — a 4-WARC test extrapolates to ~$0.20 for the full crawl slice, while the real 60-way run cost $0.49. What full fan-out buys for that premium is wall-clock: about an hour instead of about a day.

One operational note worth copying: Trafilatura(timeout=10). The FineWeb default (timeout=1) assumes dedicated cores; on shared cloud CPU at fan-out density the extraction-time tail crosses 1s en masse and each timeout costs a worker respawn. Ten seconds removes the problem at no throughput cost.

Cost receipts and reproducing the run

The cost figures are computed from the Jobs API's per-job runtime records, not estimated. cc_job_receipts.json in this repo lists every job (id, flavor, status, created_at, running_secs, cost, run label) for all 167 cpu-basic jobs I ran against this pipeline on 2026-07-14 — successes, errors, retries, cancellations, and a misconfigured first attempt included:

accounting	jobs	compute-hours	cost @ $0.01/h
the 60 completed ranks that produced this dataset	60	48.79	$0.49
the full run (incl. its errors, retries, cancels)	153	93.82	$0.94
all-in (incl. a misconfigured first attempt)	167	93.87	$0.94

To reproduce end-to-end, reproduce.py in this repo runs the identical pipeline (same datatrove commit, same crawl segment, same settings) at any scale from 4 WARCs (≈ $0.03, a smoke test) to the full 60 (≈ $0.49):

uv run https://huggingface.co/datasets/davanstrien/commoncrawl-jobs-demo/resolve/main/reproduce.py --warcs 4 --bucket hf://buckets/YOUR_USER/YOUR_BUCKET/cc-demo

(The script carries its own dependencies + Python pin via PEP 723 inline metadata — no clone, no install step.)

Afterwards reproduce.py --receipts prints your own per-job cost table from the Jobs API, so the numbers are checkable against your account rather than my word. Hardware prices: hf jobs hardware.

Contents

data/*.jsonl.gz — datatrove's standard output format: text, id (crawl UUID), metadata (url, date, language scores, filter provenance). English only (LanguageFilter default).

Licence and credit

Web text from Common Crawl, used under their terms of use; this extraction is shared under ODC-By 1.0, following the FineWeb precedent. Extraction pipeline: datatrove. Assembled by Daniel van Strien. Raw run artifacts (logs, completion markers, stats): bucket davanstrien/cc-scale-0709/stress60b/.