AxiomCodex/axiom-poi-intelligence-sample-25k

Axiom POI Intelligence (25K Sample)

Readiness Status

Public research dataset of upstream open data; see license and attribution below; no commercial or training authorization is implied.

Summary

A 25,000-row global sample of Overture Maps Foundation places (points of interest). Rows come from the pinned Overture release and are normalized to the Axiom Codex POI schema. Each row is one place. It has a stable Overture GERS identifier, a name, its primary category and category hierarchy from the Overture taxonomy, the source confidence score, point location, and an H3 resolution-8 cell for spatial joins.

The sample draws evenly from every object in the pinned Overture release (16 objects in 2026-09-23.1) and from four spread-out row groups inside each object. The result spans every inhabited continent and more than a thousand distinct primary categories. It is a static snapshot. It supports point-in-time business-composition and category analysis, not opening or closure histories.

About Axiom Codex

Axiom Codex is the governed public-data catalog maintained by Axiomancer Labs (www.axiomcodex.io). Every Axiom Codex sample is normalized to a published schema and filtered to sources whose redistribution terms are recorded in the Codex source registry. It ships with machine-readable quality evidence (quality.json, baseline.json) and a manifest that binds each file to its upstream release.

Full data

This global sample is a preview. Full U.S. Overture Places packages for all 50 states + DC, plus one consolidated all-states collection (17,288,286 rows from Overture release 2026-09-23.1, bundling the same per-state packages), are part of Axiom Codex Commercial All-access, delivered on request after purchase. See the Axiom Codex catalog. The complete California package (1,941,553 places) is a separate public research release under its upstream terms: Axiom POI Intelligence — California. This sample grants no commercial rights.

Quick Start

import json
from datasets import load_dataset

ds = load_dataset("AxiomCodex/axiom-poi-intelligence-sample-25k", split="train")
df = ds.to_pandas()
df["lat"] = df["location"].map(lambda s: json.loads(s)["lat"])
df["lng"] = df["location"].map(lambda s: json.loads(s)["lng"])
print(df["primary_type"].value_counts().head(10))

Schema

Column	Type	Description
acl_tier	string	Access control tier
confidence	float, nullable	Upstream confidence score; null when the source omits it
first_seen_at	timestamp	Date the POI was first observed
h3_index	string	H3 resolution 8 cell
id	string	Stable Overture GERS identifier
ingested_at	timestamp	Ingestion timestamp
last_seen_at	timestamp	Date the POI was last observed
location	string	JSON-encoded point ({"lat": ..., "lng": ...})
modified_at	timestamp	Last modification timestamp
name	string	Display name of the POI
normalization_version	string	Normalization standard version
observed_at	timestamp	Event observation timestamp
ontology_type	string	Ontology classification type
overture_id	string	Overture Maps Foundation identifier
primary_type	string	Primary type of the POI
record_id	string	Stable normalized source-record identifier
schema_version	string	Schema version
source	string	Primary source
source_feed	string	Exact governed registry source identifier
source_system	string	Source system
source_uri	string	Governed upstream record URI required for promotion
status	string	Operating status (open, closed); null when the source does not report one
types	string	Additional types (JSON array)

For the September 2026 Overture Places schema, primary_type and ontology_type use taxonomy.primary, falling back to basic_category only when the taxonomy primary is absent. types contains the de-duplicated taxonomy.hierarchy, taxonomy.alternates, primary type, and basic_category in that order. A missing optional source confidence remains null; present values are copied when finite and from 0 through 1. Overture operating_status maps open to open, and both temporarily_closed and permanently_closed to closed. Operating status may be unreported; a null status means the source did not report it, not that the place is open. Rows with an unrecognized reported status are dropped.

Sampling

The Overture collector reads four evenly spaced row groups from each release object (more only if those yield too few valid rows), then ranks valid overture_id values by SHA-256 across the groups read to satisfy the per-object quota. The exporter deterministically ranks record_id values by SHA-256 within source_feed. The staged manifest binds the input parquet and collector manifest.

Demo Quality Contract

This is a source_records artifact under the Axiom Codex demo quality standard (DEMO_QUALITY_STANDARD.md) poi-intelligence profile. A successful transport export is not a quality-pass claim: staging must emit quality.json and baseline.json with every critical gate passing, and only the documented Overture source is currently eligible for a governed research release.

Coverage & Caveats

• One source. Rows are filtered to source_feed=overture-maps-places. The collector's source=overture field is kept as the upstream source label.
• Unreported operating status. The source may omit operating_status; when unreported, status is null, not open. Candidate-specific missingness belongs in the release audit.
• Uneven geography. Overture coverage is denser in some regions and categories than others, and a 25,000-row global sample is thin in any one city.
• Snapshot dates only. first_seen_at, last_seen_at, and observed_at describe this snapshot. They are not business opening or closure dates.

Provenance & Source Licensing

The sample is built from the Overture Maps Foundation Places upstream collector output. The source registry documents CDLA-Permissive-2.0 with attribution and research-tier conditions; the staged manifest binds the collector receipt and input parquet digest. The projection drops per-record sources[], so per-record source licence metadata is not carried. The package ships the CDLA-Permissive-2.0 and Apache-2.0 texts and the Foursquare NOTICE covering all contributors, including AllThePlaces (CC0-1.0), with a dated change statement for modifications. The card metadata does not grant rights beyond those upstream terms.

The package ships the upstream terms and attribution files: CDLA-Permissive-2.0.txt, Apache-2.0.txt, FOURSQUARE-NOTICE.txt. Modified by Axiom Codex on 2026-10-07: the staged Overture records were normalized to the Axiom Codex POI schema and filtered to the governed field projection. This package contains a deterministic sample (25,000 of 30,000 planned rows).

Known Redactions / Exclusions

Google/Yelp, restricted raw payload, and inferred lifecycle columns are excluded by the explicit projection. The projection carries source_feed=overture-maps-places; unknown or non-research source rows fail closed. Google enrichment fields (such as google_place_id and photo_reference) are not part of the collector projection. Addresses, phone numbers, emails, websites, social links, and brand records are not projected.

The collector drops a row when its primary name contains an email address or a phone-like digit run with at least nine digits. This conservative rule can also drop benign codes such as Spbu 14 201 1144; it drops the entire row and never redacts or retains the matched name.

Some places are businesses named after a person, for example a sole proprietor. Treat names as business listings: do not use them to profile or contact individuals.

status, first_seen_at, and last_seen_at are static source observations only. They do not prove opening/closure lifecycle events. The train shard name is a Hugging Face transport convention, not a leakage-safe ML split.

The staged quality contract also fails closed on lifecycle claims: its lifecycle_config must be the reviewed not_available declaration with only a reason and the enumerated missing capabilities. Adding transitions, effective/knowledge clocks, evidence references, or adjudicated outcomes is rejected even when the declaration still says status=not_available.

ML Split

Not an ML train/test split: This dataset is provided as a single train split purely for HuggingFace dataset tooling compatibility.

Citation

Suggested source attribution and license notices for this research sample:

Hugging Face license metadata: cdla-permissive-2.0.

• overture-maps-places — CDLA-Permissive-2.0; https://docs.overturemaps.org/guides/places/. Retain per-record sources[].license; Foursquare records require Apache-2.0, the upstream NOTICE attribution, and a dated change statement when modified.

No commercial or training authorization is implied.

Contact

Maintained by Axiomancer Labs as part of Axiom Codex. For questions, corrections, or removal requests (including a place listing), open a discussion in this repository's Community tab or reach us through www.axiomcodex.io.

Governed Dataset Card Metadata

• Source / license posture: Overture Maps Places (overture-maps-places) is redistributed with package-level CDLA-Permissive-2.0 and Apache-2.0 texts and the Foursquare NOTICE covering all contributors (including AllThePlaces CC0-1.0); per-record sources[] and source licence metadata are not carried; no other source is included.
• Intended use: Research, exploration, and evaluation of static place observations, category mix, and H3-level business composition.
• Forbidden use: No profiling, tracking, or contacting of individuals named in place listings, and no treatment of status or seen-at dates as verified business lifecycle facts.
• Known missingness / bias: No authoritative lifecycle, cross-source entity resolution, evaluation, or leakage-safe ML splits; operating status may be unreported (null means unreported, not open), and coverage is uneven by region and category.
• Update cadence: Static snapshot of one pinned Overture release; later Overture releases are not reflected.
• Geographic / temporal coverage: Static place observations from the pinned Overture release; the staged manifest records the release, observation dates, and geographic scope. State variants use point-in-polygon against their pinned Census state boundaries.
• Label semantics: Place fields are static source observations, not verified entity truth or ML labels.
• Confidence / provenance method: Overture source-policy receipt, per-record source URI, and content-addressed collector and staged manifest evidence.
• Training policy: Not authorized for training under the Axiom Codex governed training-use policy; reuse rights come only from the upstream terms, and the single train split is a Hub convention, not an ML split.

Release Scope & Quality Evidence

• Scope: source_records; 25000 staged rows from upstream collector Parquet inputs.
• Observed time window: 2026-10-07T04:04:20Z to 2026-10-07T04:04:20Z.
• Machine-readable evidence: manifest.json, quality.json, and baseline.json.
• Packaging validation command: python3 tools/hf_packaging/validate.py --staged <staged-dir> --dataset poi-intelligence --release-candidate --tier research. This records the build's validation entry point and requires the Axiom Codex packaging environment, which is not included in this Hugging Face data repository. Use the quickstart above to load the data. This card does not claim a model baseline or an ML split unless a future task/evaluation configuration supplies one.
• Promotion: a release candidate requires the snapshot-bound quality report, baseline, current source-policy replay, and all applicable thresholds; blocked or unknown gates remain blocked.