DisPatch70

Aligned bike-share station status and trip records for 90+ cities, built for research on urban dispatch / rebalancing simulation. Status snapshots and trips share one directory layout, one file naming scheme, and explicit join keys.

Layout

gbfs/     <city>/<provider>/<YYYY>/<Mon>.parquet   station status, one row per station x ~15-min snapshot
tripdata/ <city>/<provider>/<YYYY>/<Mon>.parquet   trip records, one row per ride
station_dim/crosswalk.parquet                      station identity table
station_dim/snapshots/<city>/<provider>.parquet    sampled raw GBFS station_information over time

gbfs/.../2025/Mar.parquet and tripdata/.../2025/Mar.parquet cover the same city, provider and calendar month (UTC; trips are partitioned by started_at UTC to match the status feed's commit_at).

Coverage

• gbfs/: 93 city/provider pairs in 24 countries, 2023-08 (older cities) or 2024-04 (cities added later) through 2026-07, ~15-minute resolution, 349M rows. Source: the public Parquet export of MaxHalford/bike-sharing-history (git-scraped GBFS feeds), re-encoded with alignment keys added.
• tripdata/: the 12 cities whose operators publish trip records overlapping the status window: New York, Washington DC, Chicago, San Francisco Bay Area, Philadelphia, Montréal, Vancouver, Oslo, Bergen, Mexico City, Guadalajara, Buenos Aires (2023-01 → 2026-07/08 where published). Toronto publishes data but blocks non-North-American downloads; it will be added when fetched. All other systems (JCDecaux, Nextbike, Bird, Beryl, Tembici, Donkey Republic, and most municipal systems outside North America) do not publish trip-level records at all.

Alignment keys

key	gbfs/	tripdata/	note
station name	station	start_gbfs_name / end_gbfs_name	exact string match
GBFS station id	station_id	start_station_id / end_station_id	from station_dim/crosswalk.parquet; null for JCDecaux feeds (they publish no ids)
time	commit_at (UTC)	started_at / ended_at (UTC)	local wall time in started_local/ended_local + tz

start_match/end_match records how each trip endpoint was resolved: id (operator id ↔ crosswalk), name, coord (nearest station ≤ 60 m), or none. Trips with none and a null start_station_raw are dockless e-bike rides that genuinely have no station (~29% of DC and ~22% of Chicago rides); among station-referencing rides, resolution is ≥ 99.7% in every city.

gbfs/ schema

city, provider, station (name), longitude, latitude, commit_at (UTC), skipped_updates (consecutive unchanged frames before this row), bikes (available bikes), stands (available docks), station_id, capacity_ref (capacity from the crosswalk, last sampled value)

tripdata/ schema

Unified columns: city, provider, ride_id, started_at, ended_at (UTC), started_local, ended_local, tz, duration_s, start_station_raw / end_station_raw (operator's id, verbatim), start_station_name / end_station_name (as published), start_lat/lon, end_lat/lon, start_station_id / end_station_id, start_gbfs_name / end_gbfs_name, start_match / end_match, rideable_type, member_type, source_file, plus every unmapped source column verbatim as src_<original name> (string). Nothing from the original files is dropped.

Per-city id resolution: Lyft cities (NYC/DC/Chicago/SF) match start_station_raw to GBFS short_name; Oslo/Bergen/Philadelphia/Buenos Aires by digits of station_id (YOS:Station:1009 ↔ 1009, bcycle_indego_3213 ↔ 3213, 277BAEcobici ↔ 277); Guadalajara by station_id; Mexico City by digits of short_name; Montréal by name; Vancouver by the leading 0001-style code.

events/ and event_briefings/

events/           <city>/<source>/<YYYY>.parquet   one row per public event (1,032,272 rows)
event_briefings/  <city>/<source>/<YYYY>.parquet   one LLM-written briefing per event row
situations/       <city>/<YYYY>.parquet            one row per city x 30-min slot: regime, calendar, weather, air quality, notes, description

events/ unifies public-event records overlapping the status window (2023-01 → 2026-08): city permit/calendar datasets (NYC permitted events incl. street closures, Chicago park permits, SF street closures, Montréal public events, Buenos Aires mass-event permits), ESPN scoreboards for ten leagues matched to venues in the twelve cities, national public holidays, and weather / air-quality episodes derived from situations/ (source weather: rain, heavy rain, snowfall, extreme heat, severe cold, strong wind, thunderstorm, fog, poor air quality, each with start/end hour and the peak values). Columns: city, source, source_id, category (sports / concert / festival / parade / street_closure / civic / holiday / weather / other), title, description, venue_name, lat, lon, start_utc, end_utc, start_local, end_local, tz, expected_attendance, status plus every source column verbatim as src_<name>. Time precision differs by source (games to the minute, permits as start/end spans, holidays all-day, weather episodes to the hour). Coverage is uneven by design: five cities have municipal permit data, all twelve have sports, holidays and weather.

event_briefings/ turns each event row into a 1–3 sentence operational briefing for a dispatcher (where, when, how large, what it means for bike traffic), generated offline with Qwen3-14B from the prompt in the code repository (prompts/event_briefing.md). Columns: city, source, source_id, start_local, description, model, prompt_version; join back to events/ on (city, source, source_id, start_local). Descriptions use only the fields of the record; records the model failed to describe are absent rather than filled.

situations/

situations/<city>/<YYYY>.parquet   one row per city x 30-minute slot (770,688 rows, 2023-01-01 -> 2026-08-30)

What a dispatcher knows about the moment without looking at any station. Per row:

• time regime — regime ∈ night / early / am_peak / midday / pm_peak / evening on workdays, night / morning / daytime_peak / evening on weekends and public holidays, estimated per city from the tripdata layer's half-hour profile (peak = slots holding ≥ 70 % of the busiest slot's share); peak_windows spells the city's windows out; typical_trip_share (fraction of a day's trips in this slot) and typical_trips (mean trips started in this slot for this day type) quantify it. Workday peaks range from 07:30–09:00 / 15:00–17:00 in Bergen to 07:30–12:00 / 16:00–19:30 in Buenos Aires.

• calendar — weekday, is_weekend, is_holiday, holiday_name (public holidays of the city's country and state/province), calendar_note (long weekend / festive period), day_type.

• daylight — is_day, sunrise_local, sunset_local.

• weather — Open-Meteo best-match model reanalysis (1–11 km, hourly, interpolated to the slot; precipitation / rain / snowfall are the amounts falling in the slot): temperature_2m, apparent_temperature, relative_humidity_2m, precipitation, rain, snowfall, snow_depth, weather_code, weather_text, cloud_cover, wind_speed_10m, wind_gusts_10m.

• air quality — Open-Meteo CAMS model: pm2_5, pm10, us_aqi, european_aqi, ozone (model values, not monitor readings; they reproduce episodes such as the June 2023 wildfire smoke in the US Northeast but understate monitor peaks).

• notes / description — notes (JSON) lists what is special about the slot: weather episodes in progress, the holiday or calendar note, and public events active or starting within two hours (count plus up to three of the largest); notable is true when any of these is present. description is a 2–4 sentence text of the slot: routine slots are rendered from prompts/situation_template.md (description_source = template), notable slots are written by Qwen3-14B from prompts/situation_briefing.md (description_source = llm; empty until that pass has run).

Every value is for the city point (mean of its station coordinates) — the layer is the shared context of a city, not a per-station measurement. ts_local is wall-clock time in tz; ts_utc is the same instant in UTC. Built by pipeline/98_situations_build.py.

station_context/

station_context/<city>/<provider>.parquet            one row per station: OSM-derived surroundings
station_context/<city>/<provider>_profiles.parquet   one second-person self-description per station

Derived only from OpenStreetMap (Geofabrik regional extracts) plus the station crosswalk: street, neighbourhood, district, district_level (nearest named road; nearest OSM place node; the containing administrative area at one consistent admin level per city — boroughs in New York, wards/counties around Washington, community areas in Chicago, arrondissements in Montréal, bydeler in Oslo — with the finest containing area used where no single level covers a multi-jurisdiction metro), adjacent_roads (≤120 m: class, cycleway, oneway, lanes), transit (≤400 m rail/subway/tram/bus stops with operators), poi_counts (coarse categories within 300 m) and poi_tags (raw key:value counts), named_pois, buildings, landuse, neighbour_stations (≤500 m with capacity and distance). Nested fields are JSON strings. Profiles are generated with Qwen3-14B from prompts/station_profile.md in the code repository and use only the fields of the record.

station_dim/

crosswalk.parquet: one row per (city, provider, station): station_id, first_seen, last_seen, n_snapshots, name_last, names_seen (JSON list of every name observed), short_name, legacy_id, external_id, lat, lon, capacity, region_id, station_type. Built by sampling the source repo's git history on the 1st and 15th of each month (52 snapshots, 3,003 blobs).

Caveats

• The source git history is nearly empty 2024-02 → 2024-12 and 2025-07 (the archive author squashed it); station identities in those windows come from neighbouring snapshots.
• JCDecaux feeds publish no station ids; join those 22 cities by name, or by the numeric prefix many names carry ("1001 - TERREAUX").
• Vancouver trip times are rounded to the hour by the operator (privacy). 2024-02 was published as XLSX and converted to CSV before parsing.
• Trips whose local timestamp is ambiguous/nonexistent at DST transitions have started_at = NaT and were dropped (≲ 0.1%; counts in the processing logs).
• Buenos Aires 2023 has a stray index column preserved as src_ field; formats drift across years in several cities — source_file tells you which raw file each row came from.
• gbfs/ timestamps are commit times of the scraper (~every 15 min), not last_reported.

Provenance & licenses

Status: MaxHalford/bike-sharing-history (git scraping of public GBFS/JCDecaux feeds; cite the repo). Trips: the operators' open-data programs (Lyft system-data terms for NYC/DC/Chicago/SF; city open-data licenses for the rest). This dataset redistributes and re-encodes those public files; original values are preserved verbatim in *_raw, *_name and src_* columns. Situations: weather and air-quality series from Open-Meteo (CC BY 4.0; ECMWF / national weather model and CAMS data), redistributed as hourly parquet; holiday calendars via the holidays Python package.