GBFS guarantees that bike-sharing feeds are syntactically consistent — not that they are semantically comparable. This project audits 1,509 feeds across 48 countries, distils a unified taxonomy of seven data-quality classes, and releases a certified, anomaly-flagged reference dataset of 46,307 French stations so downstream consumers can filter without re-running the pipeline.
The seven data-quality classes
Five structural errors (A1–A5) that violate the implicit semantic contract of the GBFS field they populate, plus two semantic warnings (A6–A7) for spec-compliant publication patterns whose downstream interpretation is nevertheless ambiguous.
Why it matters: the Bordeaux case
The most extreme case in the French corpus is Pony Bordeaux: 2,996 entries each declaring 12 docks (a nominal 35,952 docks), while the actual mean capacity per entry is 0.03 bike/entry. After A3 reclassification, Bordeaux’s dock-based count collapses from 9,921 raw entries to 225 certified stations — a 98 % collapse, equivalent to a ×52 over-count on any supply-side metric built on the unaudited feed.
Quick start
from datasets import load_dataset
gs = load_dataset("rohanfosse/gbfs-audit-catalogue", split="train").to_pandas()
# High-confidence dock-based stations
clean = gs[(gs.station_type == "docked_bike") & (gs.audit_confidence == "high")]
print(len(clean)) # 4,721
# Operator-driven flag hotspots
gs.groupby("operator_name").agg(
n=("uid", "size"),
A3_rate=("flag_A3", "mean"),
A7_rate=("flag_A7", "mean"),
).sort_values("n", ascending=False).head(10)
No-auth alternative, straight from Zenodo:
import pandas as pd
gs = pd.read_parquet(
"https://zenodo.org/records/20125460/files/stations_gold_standard_final.parquet"
)
Resources
| Resource | Where |
|---|---|
| Interactive dashboard | gbfs-audit.streamlit.app |
| Dataset (Hugging Face) | rohanfosse/gbfs-audit-catalogue |
| Zenodo deposit (DOI) | 10.5281/zenodo.20125460 |
| Source & audit pipeline | github.com/cycling-data-lab/gbfs-audit-catalogue |
| Reproducible recipes | notebooks/catalogue_recipes.ipynb |
| Human-validation protocol | experiments/annotation/PROTOCOL.md |
| Paper (CSI 2026) | Manuscript in preparation; preprint forthcoming |
Schema — 46 columns at a glance
- 5 identifiers —
uid,station_id,system_id,system_name,source - 6 spatial / admin —
lat,lon,city,commune_name,code_commune,region_id - 4 station description —
station_name,address,capacity,n_stations_system - 13 audit pipeline —
station_type,capacity_raw,capacity_audited,flag_A1–flag_A7,operator_name,audit_confidence,fetched_at - 5 network geometry — KNN distances and density within buffers
- 2 topography —
elevation_m,topography_roughness_index - 2 cycling infrastructure —
infra_cyclable_km,infra_cyclable_pct - 1 safety —
baac_accidents_cyclistes - 2 multimodal access —
gtfs_heavy_stops_300m,gtfs_stops_within_300m_pct - 5 socio-economic context —
revenu_median_uc,gini_revenu,revenu_d1,ecart_interquar,part_menages_voit0 - 1 modal share —
part_velo_travail
FAIR by design. The catalogue ships with machine-readable descriptors — a JSON Schema, a Frictionless Data Package, a DCAT-AP record and a Croissant manifest — plus a Docker image and a 36-test CI suite for byte-level reproducibility.
Citation
@article{Fosse2026gbfs,
author = {Foss\'e, Rohan and Pallares, Ga\"el},
title = {Auditing GBFS bike-sharing feeds at country and global scale:
A reproducible anomaly taxonomy for open mobility data},
journal = {Computer Standards \& Interfaces},
year = {2026},
note = {Manuscript in preparation; preprint forthcoming}
}
@dataset{Fosse2026gbfsdata,
author = {Foss\'e, Rohan and Pallares, Ga\"el},
title = {{GBFS Audit Catalogue} v1.0},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.20125460}
}
Licence & contact
Data under ODbL v1.0 · Code under MIT.
Rohan Fossé (lead contact) · rfosse@cesi.fr · CESI École d’Ingénieurs, Montpellier, France.
Gaël Pallares · CESI LINEACT (EA 7527), Montpellier, France.
Issues and contributions on GitHub.