pretty_name: Aussie-1k
language:
- en
task_categories:
- question-answering
size_categories:
- 1K<n<10K
tags:
- benchmark
- evaluation
- australia
- australian-english
- magpie
configs:
- config_name: default
default: true
data_files:
- split: test
path: data/v6.3/test.jsonl
- config_name: v6_2
data_files:
- split: test
path: data/v6.2/test.jsonl
- config_name: v6_1
data_files:
- split: test
path: data/v6.1/test.jsonl
Aussie-1k
A benchmark of 1,500 unique questions and 1,514 reference answers for evaluating Australian knowledge and language. It covers Australian English, culture, institutions, history, Aboriginal and Torres Strait Islander topics, geography, and wildlife. The Aussie-1k name is kept for continuity.
Usage
from datasets import load_dataset
test = load_dataset("scx-admin/aussie-1k", split="test") # 1,500 questions
Present only question to the model. Keep answers and curation metadata hidden. Score each question once, with equal weight, using a documented correctness rubric that allows partial credit. Exact match and token F1 can supplement semantic scoring. Record the dataset commit, model, prompt, settings, scoring method, and tool access.
Composition
| Measure | Count |
|---|---|
| Questions | 1,500 |
| Reference answers | 1,514 |
| Questions with one reference | 1,486 |
| Questions with two references | 14 |
| Training or validation splits | 0 |
| Coverage area | Examples |
|---|---|
| Australian English | Spelling, slang, idioms, local terminology |
| Culture, media, and sport | Entertainment, literature, music, food and customs, sport |
| Civic knowledge | Institutions, Medicare, superannuation |
| History and Indigenous topics | Australian history, Aboriginal and Torres Strait Islander peoples and cultures |
| Geography | States, territories, towns, landmarks |
| Wildlife and safety | Animals, beaches, bushwalking |
The original 984 questions have no topic labels, so their category proportions are unknown. The 516 later additions carry topic labels in their metadata.
Metadata and schema
| Property | Value |
|---|---|
| Project | SouthernCrossAI / SCX.ai MAGPiE evaluation |
| Language | Australian English (en-AU) |
| Format | UTF-8 JSON Lines |
| Split | test |
| License | Unspecified |
| Field | Type | Description |
|---|---|---|
id |
string | aussieqa- plus the first 20 hexadecimal characters of the question's SHA-256 hash |
question |
string | Input question |
answers |
list of strings | One or two reference answers |
Questions are normalised with Unicode NFC and trimmed before hashing and duplicate grouping. Distinct answer variants are retained. IDs depend on question text, not row order or answer wording.
Provenance and limitations
The benchmark originated in the SCX.ai MAGPiE evaluation project. Original curation metadata records difficulty, obscurity, quality, decisions, and source-row mappings; these are dataset ratings, not model scores.
The original references date from November 2025 and have not been fact-checked again. Original item-level sources and reuse rights are not established. Exact duplicate removal does not eliminate paraphrases, and alternate references may differ in precision. The dataset has no documented representative sampling frame or validated pass threshold.
Version history
| Version | Changes |
|---|---|
| 0.1 | Original 1,000-row curated dataset. |
| 0.2 | Merged duplicate questions: 984 questions, 998 references; added stable IDs, a test split, and separate curation metadata. |
| 0.3 | Added 16 source-checked questions: 1,000 unique questions and 1,014 references. |
| 0.4 | Added 500 source-linked, multi-model-reviewed questions: 1,500 unique questions and 1,514 references. |
The 16 additions were authored by OpenAI Codex on 14 October 2026 and fixed before model evaluation. They cover Australian English, culture/media/sport, civics, and history/Indigenous topics (3 each), plus geography and wildlife (2 each). Difficulty estimates at levels 1–4 are 2/7/6/1. This distribution is editorial; the original 984 questions and references are unchanged.
Files: original source, curation audit, additions and sources, addition metadata, and release manifest. Historical filenames retain their original identifiers.
The 500-question supplement (0.4)
Claude Opus 5.5 (Anthropic) authored the 500 questions and one-sentence references on 14 October 2026, working as 11 parallel authoring agents with one topic slice each. Each agent was given the existing questions to avoid repeats and an item-level source requirement. The supplement was frozen before any model was evaluated on it. The 0.3 questions and references are unchanged and form the first 1,000 rows.
| Area | Questions | Slices |
|---|---|---|
| Australian English | 90 | slang and vernacular; spelling, usage and regional vocabulary |
| Culture, media, and sport | 135 | food, drink and customs; arts and media; sport |
| Civic knowledge and everyday systems | 90 | civics and law; tax, health, work, education, housing and transport |
| History and Indigenous topics | 90 | national history; Aboriginal and Torres Strait Islander peoples, cultures and history |
| Geography and travel | 50 | places, regions, landmarks, time zones and external territories |
| Wildlife and safety | 45 | fauna, flora, environment, bushfire, beach and first-aid guidance |
Difficulty estimates at levels 1–4 are 69/219/166/46, close to the original curation distribution. Questions avoid current office-holders, changeable amounts and "latest" results.
Review. Two SCX.ai-hosted models, MiniMax-M3 and GLM-5.3-Flash, independently rated each candidate for Australian fit (1–5), reference correctness, ambiguity and time-stability. A calibration run accepted all 16 version 0.3 additions and rejected four planted faults. Candidates passed only if both judges accepted them with no flag. Claude adjudicated every flagged item against sources: it was kept, revised and re-reviewed by both judges, or dropped. Judges were sometimes wrong, and those objections were overruled with recorded evidence. Claude then read every remaining candidate and removed overlapping, answer-leaking or weakly attested items. Near-duplicates against earlier questions and within the supplement were flagged with E5-Mistral-7B-Instruct embeddings (cosine ≥ 0.82) and resolved by hand.
Of 608 drafts, 26 were dropped and 500 were selected: 45 per slice (50 for geography), keeping each slice's difficulty mix and ranking by the judges' summed fit scores. No model answers were used in selection. Of those selected, 479 passed both judges unchanged, 10 were revised then passed, and 11 were kept after adjudication. Per-item topic, difficulty, source URL, both judges' ratings and adjudication notes are in addition metadata.
Limitations. The questions are model-authored and the judges are models; no independent human expert review has been done. Of the 500 source links, 331 are Wikipedia, because the authoring agents' web-search allowance ran out and later facts were checked by fetching pages directly. A few sources are JavaScript-rendered or block automated access, so they could not be machine-verified. Australian first-aid answers follow Australian guidance current in October 2026 and are evaluation material, not medical advice.
Files: additions and sources, addition metadata, release manifest and build script.