TenantBench

TenantBench is a decision-derived benchmark for evaluating whether language models can help self-represented tenants prepare for Australian residential tenancy disputes from a de-identified lay narrative and a jurisdiction hint. This repository contains the dataset-only v6.2 release; code, frozen scoring artifacts, and the retriever checkpoint are distributed separately.

TenantBench separates six capabilities:

• D1 — Applicable Law: provisions applied by the tribunal.
• D2 — Evidence Held: evidence types already available to the tenant.
• D3 — Missing Evidence: evidence whose absence materially weakened an issue.
• D4 — Evidence-Supported Law: D1 provisions retained when the decision establishes a grounded provision–evidence connection.
• D5 — Decision-Relevant Evidence: evidence that materially affected the tribunal's finding under each D4 provision.
• D6 — Evidence Direction: whether that evidence-supported determination was pro-tenant, anti-tenant, or neutral.

Configurations and splits

Configuration	Split	Cases	Intended use
gold	dev	53	Development and selection
gold	test_iid	98	In-distribution evaluation
gold	test_ood_sa	165	Held-out South Australian evaluation
silver	original_pool	473	Training and retrieval
silver	pipeline_refined	1,756	Training and retrieval
audit_quarantine	quarantine	995	Audit only; never active training or evaluation data

The active full-schema corpus contains 316 gold and 2,229 silver decisions. Gold cases received item-level dual review and housing-expert final judgment. Silver cases follow the model-based construction and refinement workflow but do not receive the gold tier's independent dual-human review and expert final judgment.

Loading

from datasets import load_dataset

gold = load_dataset("Essie0715/TenantBench", "gold")
silver = load_dataset("Essie0715/TenantBench", "silver")

test_iid = load_dataset(
    "Essie0715/TenantBench",
    "gold",
    split="test_iid",
)

For a private repository, authenticate first with hf auth login or pass an authorized token through the standard Hugging Face client.

Task input boundary

The intended model input consists only of:

• input.lay_narrative
• input.jurisdiction_hint

input.evidence_the_tenant_has_or_can_get is reference metadata and must not be supplied as an additional task input. Quarantined records must not be used as evaluation gold or active training data.

Schema and provenance

See SCHEMA.md for field definitions and DATASHEET.md for collection, annotation, de-identification, limitations, and provenance details. The auxiliary d3_reference.json file contains the fixed evidence-type reference used by the release.

The configured Parquet files are a schema-stable projection for the Hugging Face Dataset Viewer: fields absent from a particular source record are represented as null, while every existing value is retained. Exact JSONL copies of the GitHub v6.2 dataset files use the raw_*.jsonl filename prefix. Projection details are recorded in PROJECTION.json.

The source corpus comprises published written tribunal decisions from New South Wales, Victoria, and South Australia. It does not represent the full universe of tenancy applications. Decision-derived supervision also captures what the written judgment records, not every item a tenant could have prepared. This limitation is especially important for D3 missing-evidence targets.

Privacy and responsible use

Public-facing records are de-identified derivatives of published decisions. Names and street-level addresses are generalized where identified, while legally material dates, amounts, jurisdictions, and section numbers are retained. Source citations and URLs are preserved for provenance, so motivated users may still locate the underlying public judgment.

TenantBench is intended for legal-NLP benchmarking, retrieval research, evidence reasoning, and access-to-justice research. It is not legal advice and is not intended to support decisions about real individuals.

Licensing

Authored annotations, rewritten narratives, and derived dataset structure are licensed under CC BY-NC 4.0. Short statutory excerpts and decision-derived snippets remain third-party source material and are not relicensed by the dataset licence. Full tribunal decisions are not redistributed. See LICENSE-DATA.txt and THIRD_PARTY_NOTICES.md.

Citation

The citation remains anonymized during review. See CITATION.bib; final author and venue metadata will be added after the anonymous-review period.