scx-admin/au-building-legislation-small

Australian Building Legislation Small

The full benchmark contains all 500 examples in a single test split. It tests selection of applicable Australian building legislation from synthetic applications across all eight states and territories, using NCC 2022, NCC 2025 and state provisions. Version 1.0 was frozen on 11 October 2026, with a legal reference date of 10 October 2026.

499 records have settled benchmark labels. The remaining record, bls-v1-0006, is included in the full benchmark with scored=false: its pool energy requirements cannot be determined reliably from the stated equipment scope. Its stored labels are provisional and must not be treated as settled gold. There are also 39 null-gold questions among the 499 scored records. These are excluded from the 7,194-question scoring denominator.

Load all 500 examples

import json
from datasets import load_dataset

benchmark = load_dataset("scx-admin/au-building-legislation-small", split="test")
assert len(benchmark) == 500

for row in benchmark:
    request = json.loads(row["request_json"])  # Model input: state and questions
    record = json.loads(row["record_json"])    # Full lossless record, including labels
    if row["scored"]:
        gold = json.loads(row["gold_json"])
        # Score only non-null gold values; see docs/LABEL_SPEC.md.

The full benchmark combines the original train, dev and test allocations. It is an evaluation collection of all 500 examples, not an independently held-out 500-example test set. original_split retains those allocations (250 train, 62 dev, 188 test including the contested record). For the original scored partitions:

scored = load_dataset("scx-admin/au-building-legislation-small", "scored")
assert {split: len(rows) for split, rows in scored.items()} == {
    "train": 250, "validation": 62, "test": 187
}

Format

The Parquet columns are id, dataset_version, as_of, original_split, family_id, difficulty, jurisdiction, state_format, scored, request_json, gold_json and record_json. JSON strings preserve the original heterogeneous state (text or object), typed question maps and labels without Arrow type coercion. record_json includes the complete original record: request, facts, gold, rationale, citations, candidate mapping, hard-negative tags, token counts, flags and provenance. Only request_json is model input; other columns can disclose labels or generation facts.

data/benchmark.jsonl provides all 500 full records as ordinary JSONL. The original release files are preserved byte for byte, including the 499-record all.jsonl, the separate one-record contested.jsonl, request-only split exports and the frozen manifest. The manifest and its original filenames refer to that frozen release, while publication.json describes this Hugging Face packaging. The publication adds no labels and does not relabel the contested example.

Documentation and source references

• Detailed dataset card: composition, quality checks, agreement, results and limitations.
• Label specification: question types, applicability rules and null-gold scoring.
• Method: generation, blind checks, adjudication and evaluation.
• Annotator handbook: instructions used in independent annotation.
• Adoption review and guide corrections: source interpretation and documented repairs.
• Unit catalog: candidate keys, source UIDs, short excerpts and source identifiers; catalog provenance.
• NCC edition adoption rules: the frozen edition-selection rules referenced by the labels (reference source, not a packaged evaluation runner).

Source text retains its source-specific terms. See LICENSE.md and SOURCE_ATTRIBUTION.md for copyright and attribution. Australian Standards are referenced by designation, not reproduced. Applications and associated contact strings and copyright boilerplate are synthetic and imply no real-world affiliation. Historical audit paths and model/annotation provenance are retained in the frozen artifacts for reproducibility.

Existing evaluation results

On the original 187-example held-out test partition (2,870 scored questions):

System	All-question accuracy
Jev 1.13	74.74%
BM25 + train-majority baseline	51.22%
Tev1-0.8B experimental	43.03%
some_model	37.07%

See the held-out report and full 499-scored-example report for metric definitions, coverage and slices. These are existing evaluation results, not new runs for this publication. Jev used OpenRouter; the two local models ran on a 6 GB GPU with an 8,192-token input limit. Missing answers and execution failures count as incorrect. Neither Clef nor Clef-Flash was evaluated.

The dataset is synthetic, small and deliberately rich in hard cases. It is not a representative sample of permit applications and should not be used as a basis for live legal or building certification decisions. Reported local model scores include context-limit and GPU memory failures. Generation and first annotation used mixed model provenance, with an independent GLM second opinion; see the detailed card for limitations and all disclosed review procedures.