README.md

---
pretty_name: LOCUS v1.0
language:
- en
task_categories:
- text-classification
size_categories:
- 1M<n<10M
tags:
- law
- legal-nlp
- local-government
- municipal-law
- ordinances
- classification
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: header
    dtype: string
  - name: content
    dtype: string
  - name: is_substantive
    dtype: bool
  - name: function
    dtype: string
  - name: topic
    dtype: string
  - name: source_jurisdiction_type
    dtype: string
  - name: state
    dtype: string
  - name: city
    dtype: string
  - name: county
    dtype: string
  - name: enforcement_discretion
    dtype: float64
  - name: opacity
    dtype: float64
  - name: paternalism
    dtype: float64
  - name: problem_salience
    dtype: float64
  splits:
  - name: train
    num_examples: 2211516
rai:dataLimitations: >
  - Coverage is not exhaustive. LOCUS v1.0 does not represent every U.S.
  municipal or county jurisdiction, and jurisdictions vary in how completely
  their local law is digitized and codified.

  - The `function` and `topic` labels are model-assigned and have not been fully
  human-validated. Label noise should be expected, particularly at taxonomy
  boundaries (e.g., Rules vs. Process, Buildings vs. Zoning).

  - The topic taxonomy is intentionally coarse (Buildings, Business, Nuisance,
  Zoning, Other) and does not capture fine-grained subject matter important for
  some downstream tasks.
  
  - Local law changes over time. The dataset reflects a snapshot and may not
  match the current law of any given jurisdiction.
rai:dataBiases: >
  - Geographic and population skew: larger and more urbanized jurisdictions, and
  jurisdictions that publish their codes through major codification vendors, are
  likely overrepresented relative to small or rural municipalities.

  - State-level legal regimes differ, so the distribution of functions and
  topics is not uniform across states; aggregate label frequencies should not be
  assumed to generalize to any individual jurisdiction.

  - The taxonomy itself encodes choices about what counts as "substantive" law
  (Rules and Enforcement) versus procedural, contextual, or structural
  provisions. Alternative legal-theoretic framings would produce different
  labels.

  - Labels were produced by automated classifiers and may reflect biases of
  those classifiers, including biases inherited from their training data.

  - English-language only. Non-English text occasionally present in U.S. local
  codes is not handled as a separate class.
rai:personalSensitiveInformation: >
  LOCUS v1.0 is built from publicly enacted municipal and county law, which is a
  matter of public record. However, local ordinances can incidentally contain
  references to identifiable individuals (e.g., named officials, sponsors, or
  parties to specific proceedings) and to specific properties or businesses
  (e.g., addresses in zoning provisions, license holders, named establishments).
  No effort has been made to redact such incidental references. The dataset does
  not contain non-public personal data, biometric data, health data, financial
  account information, or government-issued identifiers.
rai:dataUseCases: >
  Intended uses include legal text classification research, analysis of
  local-law structure and composition, substantive vs. non-substantive filtering
  for downstream legal NLP pipelines, refinement of legal taxonomies, and weakly
  supervised or human-in-the-loop annotation workflows. The dataset is not
  appropriate as legal advice, as a substitute for human legal review, as a
  complete census of U.S. local law, or as a fully validated benchmark in the
  absence of additional human auditing.
rai:dataSocialImpact: >
  Anticipated positive impacts include lowering the cost of empirical research
  on local government law, supporting tools that help residents, journalists,
  and researchers understand municipal regulation, and enabling reproducible
  study of how local jurisdictions structure their codes. Anticipated risks
  include over-reliance on automated classifications in settings with legal
  consequences, misinterpretation of aggregated label frequencies as
  authoritative statements about the law, and downstream tools surfacing model
  outputs to end users as if they were vetted legal information. Users deploying
  systems built on this dataset should clearly disclose the automated and
  unaudited nature of the labels and should not present outputs as legal advice.
rai:hasSyntheticData: false
license: cc-by-nc-4.0
---

# LOCUS v1.0

## Dataset Summary
LOCUS v1.0 is a chunk-level dataset of U.S. municipal and county law text labeled by legal function. Each eligible chunk is assigned a `function`, a binary `is_substantive` label, and all substantive provisions are assigned a `topic`.
The dataset is intended for legal text research, local-law structure analysis, substantive filtering, and downstream taxonomy refinement.

## Dataset Structure
The unit of analysis is a text chunk derived from local law documents.

## Scale
Approximate scope for v1: `2,211,516` ordinances

## Label Schema

### Function
Allowed values:
- `Context`
- `Rules`
- `Process`
- `Enforcement`

### Substantive Indicator
The released dataset enforces the following deterministic rule:
- `is_substantive = 1` for `Rules` and `Enforcement`
- `is_substantive = 0` for `Context`, `Process`, and `Structural`

### Topic
Used only when `is_substantive = 1`.

Allowed values:
- `Buildings`
- `Business`
- `Nuisance`
- `Zoning`
- `Other`

## Recommended Uses
LOCUS v1.0 is appropriate for:
- legal text classification research
- local law structure analysis
- substantive versus non-substantive filtering
- downstream taxonomy refinement
- weakly supervised or human-in-the-loop legal NLP workflows

## Out-of-Scope Uses
LOCUS v1.0 should not be treated as:
- legal advice
- a substitute for human legal review
- a complete census of all U.S. local law
- a fully human-validated benchmark without additional auditing

## Citation