AndreyKopanev/f1-data-mining

Formula 1 research archive, 2014–2025

A versioned assembly of historical Formula 1 data from Jolpica and FastF1, collected September 10–11, 2026. This is an independent research archive, not an official Formula 1 product. The files preserve broad inputs before modeling or lap filtering. Collection and analysis code are maintained in a separate private team repository.

Downloads

• —analysis.tar.gz: Parquet tables, session metadata, original collection receipts, coverage reports, and Italy 2018 recovery evidence.
• —reproduction.tar.gz: raw source responses, FastF1 caches, and historical collection status. Add this to the analysis archive for the full archive.
• —manifest.json: SHA-256 checksums, byte sizes, file counts, collection coverage, and the packaging source commit.

Each archive extracts directly into a data root: tables/, receipts/, etc. Use a specific repository commit as the revision for reproducible downloads. No account is required to download. Archives preserve heterogeneous table schemas; they are not a single Hugging Face load_dataset() table, and the dataset viewer is not configured.

uv tool install huggingface-hub
hf download AndreyKopanev/f1-data-mining analysis.tar.gz manifest.json \
  --repo-type dataset --revision COMMIT_SHA --local-dir download

Verify the archive SHA-256 against manifest.json before extracting into a new empty directory. The private team's f1-data download command automates pinned downloads, checksum checks, and safe extraction. The manifest's code commit identifies the packaging implementation, not the original collector launch.

Sources and coverage

• —Jolpica: 2014–2025 calendars, circuit information, drivers, constructors, race and qualifying results, sprints, pit stops, and 2014–2017 race laps.
• —FastF1, version 3.8.3: 2018–2025 practice, qualifying, sprint and race sessions, including lap/sector timing, tyres, weather, results, track/session status, and race-control messages.

858 running sessions are collected; five source-confirmed cancellations account for all 863 scheduled sessions in the detailed period. See coverage.json and coverage.md. The full uncompressed archive is approximately 3.9 GiB; most of it is raw responses and caches. High-frequency car/position telemetry, preseason testing and the incomplete 2026 season are excluded.

Organization and interpretation

tables/jolpica/season=YYYY/ contains flattened Parquet tables; nested source fields use dotted column names. tables/fastf1/season=YYYY/round=RR/session_name/ contains separate table families plus session metadata. Use season and round for event joins, retain source driver IDs, and treat driver numbers as session-scoped. Do not concatenate unrelated table families or count their combined rows as independent observations.

Lap tables retain invalid/deleted/interrupted/pit laps and quality flags. Classifications may reflect later penalties; they are not point-in-time forecasts. FastF1 reconstructs some values and uses sources including Jolpica, so the two sources are not independent. Timing durations are usually relative to the timing stream origin, not necessarily the scheduled session start.

Italy 2018 race recovery leaves Ricciardo tyre fields missing and marks TyreDataUncertain=True; it does not invent tyre histories. Fuel loads, setup, damage, upgrades and actual tyre condition cannot be reconstructed here.

Original receipts retain collector-machine paths for provenance. Their tables/ suffix identifies the corresponding output in a downloaded archive. Archived worker status describes the old collection, not a current process.

Attribution and caches

Underlying data rights remain with the respective sources and rights holders; no new permissive license is asserted over the assembled source data. The source software licenses do not themselves license all upstream data.

The optional FastF1 caches contain Python pickles. Only load caches from a trusted publisher after verification. Ordinary analysis can read the Parquet files without loading these caches.