LOCUS-v1 / README.md
ChristopherVu's picture
LOCUS v1.1 with dimension scores
88cb1e1 verified
|
Raw History Blame
12.3 kB
metadata
pretty_name: LOCUS v1.1
language:
  - en
license: cc-by-nc-4.0
size_categories:
  - 1M<n<10M
task_categories:
  - text-classification
tags:
  - law
  - municipal-law
  - model-annotations
configs:
  - config_name: sections
    default: true
    data_files:
      - split: train
        path: sections/*.parquet
  - config_name: labels
    data_files:
      - split: train
        path: labels/*.parquet
  - config_name: human_audit
    data_files:
      - split: train
        path: human_audit/*.parquet
  - config_name: counties
    data_files:
      - split: train
        path: counties/*.parquet

LOCUS v1.1

LOCUS (Local Ordinance Corpus for the United States) contains municipal and county legal text harmonized to county geography. This release includes three models' annotations and a human audit for primary function, regulated topic where applicable, and rubric confidence.

Paper · Code

Annotations use Gemini (gemini-3.8-flash), OpenAI (gpt-6-luna, low reasoning), and Jev (jev-1.14.0).

Subsets and coverage

Subset Rows Contents
sections 3,073,770 Default: identifiers, jurisdiction, text, consensus, all model labels, and dimension scores (29 columns).
labels 3,073,770 The section identifiers and label columns, without jurisdiction or legal text (17 columns).
human_audit 250 Human consensus, independent eligibility flags, and the three models' labels.
counties 3,144 The county harmonization universe, including counties without a selected document.

The sections contain 2,360 distinct documents. The county table contains 3,144 counties: 2,410 represented (76.7%) and 734 without a selected document. These are counts of the released county selection, not a claim that every ordinance in a represented county is included.

The sections and labels rows share the same order, sorted by state, document_id, and section_index. The composite key (document_id, section_index) is unique. Parquet files use Zstandard compression.

Loading

from datasets import load_dataset

repo_id = "LocalLaws/LOCUS-v1"
sections = load_dataset(repo_id, "sections", split="train", streaming=True)
labels = load_dataset(repo_id, "labels", split="train", streaming=True)
human_audit = load_dataset(repo_id, "human_audit", split="train")
counties = load_dataset(repo_id, "counties", split="train")

Streaming avoids downloading and materializing the entire sections table before iteration.

Column glossary

sections

Column Type Nullable Meaning
document_id string no Stable source document identifier; join key with section_index.
section_index int64 no Original index within its document; gaps and an initial zero are allowed.
state string no Uppercase two-letter jurisdiction state abbreviation.
jurisdiction_name string no Selected jurisdiction name from county harmonization.
jurisdiction_type string no city, county, city-county, or other; city-county combines municipal/county governments or functions, and other is Washington, DC.
jurisdiction_geoid string yes Single 5-digit county, 7-digit place, or 10-digit county-subdivision GEOID; null when several apply. Leading zeros are retained.
represented_county_fips list<string> no Nonempty list of represented county FIPS codes sorted by FIPS.
represented_county_name list<string> no County names aligned with represented_county_fips.
header string yes Source header with leading Markdown # marks and surrounding whitespace removed.
content string no Unchanged source section content.
function_majority string yes Function receiving at least two matching model votes; otherwise null.
function_votes int8 no Largest number of matching function votes, from 0 to 3.
topic_majority string yes Topic receiving at least two matching votes; null also represents a not-applicable majority.
topic_votes int8 no Largest number of matching topic votes, including not applicable, from 0 to 3.
gemini_function string yes Gemini primary function; null for an unavailable answer.
gemini_topic string yes Gemini topic (the codebook subcategory); null when inapplicable or unavailable.
gemini_confidence string yes Gemini rubric confidence: Low, Medium, or High; null when unavailable.
openai_function string yes OpenAI primary function; null for an unavailable answer.
openai_topic string yes OpenAI topic (the codebook subcategory); null when inapplicable or unavailable.
openai_confidence string yes OpenAI rubric confidence: Low, Medium, or High; null when unavailable.
jev_function string yes Jev primary function; null for an unavailable answer.
jev_topic string yes Jev topic (the codebook subcategory); null when inapplicable or unavailable.
jev_confidence string yes Jev rubric confidence: Low, Medium, or High; null when unavailable.
jev_function_probs struct<float32> yes Jev scores keyed by Context, Rules, Process, Enforcement, Structural; null iff jev_function is null.
jev_topic_probs struct<float32> yes Jev scores keyed by Buildings, Business, Zoning, Nuisance, Other; null iff jev_topic is null.
enforcement_discretion float64 yes Enforcement discretion score, standardized over non-Structural sections; null iff function_majority is Structural.
opacity float64 yes Opacity score, standardized over non-Structural sections; null iff function_majority is Structural.
paternalism float64 yes Paternalism score, standardized over non-Structural sections; null iff function_majority is Structural.
problem_salience float64 yes Problem salience score, standardized over non-Structural sections; null iff function_majority is Structural.

labels

document_id, section_index, the four consensus columns, the nine per-model label columns, and the two Jev probability columns have the same definitions and types as sections.

human_audit

Column Type Meaning
audit_id string Stable audit excerpt identifier.
text string Original human-audit excerpt text.
human_function, human_topic, human_confidence nullable string Human majority consensus, recased to the release vocabularies.
human_function_eligible, human_topic_eligible, human_confidence_eligible bool Independent flags identifying fields eligible for evaluation.
human_annotations int64 Number of human annotations for the excerpt.
gemini_*, openai_*, jev_*, jev_*_probs as in sections Model labels and scores joined by audit_id.

counties

Column Type Meaning
county_fips string Five-digit county FIPS code, preserving leading zeros.
county_name, state string County name and uppercase state abbreviation from harmonization.
document_id nullable string Selected representative document; null for an uncovered county.
jurisdiction_name, jurisdiction_type nullable string The selected jurisdiction, with the same definitions as sections.
selection_reason string Recorded harmonization selection reason.
section_count nullable int64 Released sections in the selected document; null when none is selected.

Pre-cleaning page counts are excluded because they do not describe the released text.

Labels, consensus, and missing values

topic is the annotation codebook's subcategory. Function values are Context, Rules, Process, Enforcement, and Structural. Topic values are Buildings, Business, Zoning, Nuisance, and Other. Confidence values are Low, Medium, and High.

Unavailable or unsuccessful model answers use real null values for all of that model's labels. A null topic with a non-null function means topic does not apply (Context, Process, or Structural). Topic is non-null exactly for answering models whose function is Rules or Enforcement. No ERROR strings or inferred labels are inserted. Per-model missing counts are in manifest.json.

Section rows without a model answer: Gemini 60; OpenAI 1; Jev 0.

Consensus counts only models with valid answers. An answering model's inapplicable topic counts as a vote for not applicable; a missing model contributes no vote. A majority requires at least two matching votes. topic_majority null with topic_votes at least 2 means a majority for not applicable; at most 1 means no majority. All three agree when both vote counts equal 3. Confidence is reported separately for each model.

Human-audit agreement

The released audit contains 250 excerpts. Each excerpt has three human annotations.

The following table reports excerpts with human majority consensus for each field. These are consensus eligibility rates, not pairwise or chance-corrected inter-annotator agreement statistics.

Human field Eligible consensus / audit excerpts
function 226/250 (90.4%)
topic 216/250 (86.4%)
confidence 244/250 (97.6%)

Human function, topic, and confidence have separate eligibility flags and must be evaluated independently. An eligible blank human topic is a valid null; an ineligible blank is unscored. Joint evaluations require all relevant eligibility flags. Shared audit_id values should remain together when splitting related human-label data.

The next table reports exact model agreement with eligible human consensus. A row is scored only when that field is human-eligible and the model returned a valid function. Null topics from answering models are valid not-applicable answers and are compared as such. Missing model answers reduce scored coverage, and do not enter the agreement denominator.

Model Field Exact matches / scored Scored / eligible
Gemini function 193/226 (85.4%) 226/226 (100.0%)
Gemini topic 185/216 (85.6%) 216/216 (100.0%)
Gemini confidence 141/244 (57.8%) 244/244 (100.0%)
OpenAI function 177/226 (78.3%) 226/226 (100.0%)
OpenAI topic 172/216 (79.6%) 216/216 (100.0%)
OpenAI confidence 108/244 (44.3%) 244/244 (100.0%)
Jev function 183/226 (81.0%) 226/226 (100.0%)
Jev topic 164/216 (75.9%) 216/216 (100.0%)
Jev confidence 113/244 (46.3%) 244/244 (100.0%)

Collection and limitations

The providers use the shared codebook and identical header-plus-content input, capped at 16,000 characters. Truncation counts are measured during the build and recorded in manifest.json.

Only Jev supplies probability scores. They are preserved as float32 without normalization; provider rounding means sums may differ from 1. They are provider scores, not verified calibrated probabilities. Evaluate calibration on the human audit before interpreting them as calibrated probabilities.

Confidence follows a rubric with Low precedence. Compare confidence within each source; the rubric can produce more Low responses than the human audit. Confidence agreement measures rubric-label matching, not calibration.

The shared Tennessee Mayor-Aldermanic model charter has several place GEOIDs and therefore a null jurisdiction_geoid. Represented counties come exclusively from county harmonization. Washington, DC has jurisdiction_type other. Uncovered counties retain null document and jurisdiction fields.

Provenance

manifest.json records input hashes, exact model and prompt metadata, measured counts, missing answers, source-code hashes, and deterministic checks against run ledgers. Unavailable actual run dates remain null with an explanation; annotation collection timestamps are not substituted for run dates.

License

Declared dataset license: cc-by-nc-4.0.

Citation

@article{peskoff2026freeing,
  title={Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States},
  author={Peskoff, Denis and Barrow, Joe and Vu, Christopher and Davenport, Diag},
  journal={arXiv preprint arXiv:2606.19334},
  year={2026}
}