lfqian/Annulus-Economic-Benchmark

Annulus Economic Benchmark

Annulus Economic Benchmark contains three tasks built from economic statistical series: self-reported knowledge boundaries, historical value recall, and retrospective monthly forecasts. All model-facing prompts are in English. This release contains 8,190 fact records and 26,460 questions.

Coverage

Component	Target periods	Facts per period	Task 1 per period	Task 2 per period	Task 3 per period
Annual	1990–2024, inclusive	180	360	180	—
Monthly	January 2025–September 2026, inclusive	90	180	90	90

The annual component contains 6,300 facts and 18,900 questions. The monthly component contains 1,890 facts and 7,560 questions. Across both components, Task 1 has 16,380 questions, Task 2 has 8,190, and Task 3 has 1,890. The annual and monthly panels differ; within a monthly target period, the three tasks use the same country–indicator objects.

Tasks

Task 1 examines whether a model claims to know an outcome or treats it as a forecast. Task 2 checks recall of the recorded value. Task 3 elicits probabilities for the next month's direction of change.

Task 1 — Self-reported knowledge boundary. Each fact produces a recall question and a prediction question. The choices distinguish factual answers, inability to recall an outcome outside the model’s knowledge, and prediction rather than claimed prior knowledge. Gold labels are conditional on the inside or outside boundary condition. A reference month, year, or advertised model cutoff does not by itself establish whether a particular fact is inside a model’s knowledge. Users must specify their boundary-assignment protocol when reporting Task 1 scores.

Task 2 — Historical value recall. Each question asks for one economic indicator's value for a specified annual or monthly reference period, with four numerical answer choices. The model selects the recorded value, rounded to two decimal places as stated in the question. Task 2 contains 8,190 ordinary factual-recall questions: 180 per annual target year and 90 per target month from 2025 onward. The reference period does not establish the value's first-publication date or the complete information set available at that time.

Task 3 — Monthly forecast. The forecast date is the end of the preceding month. The model predicts whether the target month’s statistical value will increase, remain unchanged, or decrease relative to the preceding month. It must return a probability for every labelled choice, summing to one, and a short reason. These are retrospective forecast questions with recorded outcomes, not a claim that the historical information available at each forecast date has been fully reconstructed.

Sources and construction

The data are constructed from frozen statistical-table downloads from the Bank for International Settlements, World Bank WDI and GEM, OECD, and national statistical or central-bank publications. The release includes 50 verified supplemental observations, with source provenance and any quotation conversions retained in the evidence. Supplements include explicitly identified attributed sources where an official original was inaccessible. Source values are used to construct numerical choices, factual-recall answers, and forecast outcomes.

Five topics were replaced only in September 2026 to use available published outcomes while preserving the regional counts: Japan CPI → Japan long-term government bond yield; UK CPI → UK long-term government bond yield; South Africa CPI → Kenya policy rate; Morocco CPI → Nigeria policy rate; and Algeria CPI → Ghana policy rate. These substitutions are recorded in the release metadata and should be considered when comparing September with earlier months.

Two-column evaluation table

Select the task and target year together under Subset, for example task1_1990, task2_1990 or task3_2026. Split is always test. Every displayed row has exactly two top-level columns:

Column	Content	Use
input	The complete prompt as a plain text string, including instructions, the question and labelled options.	Pass this string directly to the model.
gold	The reference answer for that same row.	Read only for offline scoring.

Task 1's gold is an object with the existing inside and outside option letters. Select the condition according to your boundary-evaluation protocol; the release does not guess a model-specific knowledge boundary from a date. Task 2's gold is the option letter for the correct numerical value. Task 3's gold is the observed outcome's option letter; its input still requests a probability distribution and a reason. Task 3's mapping from option letters to directions is retained in each input, so a particular letter must not be assumed always to mean an increase or decrease.

Task 1 and Task 2 each have 37 task/year configurations, covering 1990–2026. Task 3 has task3_2025 and task3_2026. There are 76 configurations in total, each with one test split. Years are configurations rather than split names so that the Dataset Viewer does not encounter dozens of nonstandard splits. The month remains written in each monthly question.

Validation and limitations

Numerical values, derived answers, and export consistency have been checked against the frozen source records and documented supplements. The release has not undergone a complete item-by-item semantic review or a full LLM rewriting pass; the questions are generated deterministically from the statistical records.

The source snapshots may contain revisions and are not historical real-time vintages. First-publication dates have not been comprehensively verified. Reference-period labels and evaluation dates must not be treated as public-release dates: annual series may follow fiscal-year reporting conventions, and a December 31 grouping date does not establish a release date. Monthly interest-rate observations are not all month-end rates; each series retains its stated measurement basis. Supplemented exchange-rate observations retain their quotation basis, transformations, and effective precision.

The dataset provides prompts, outcomes, and provenance for evaluation. This release does not include model-evaluation results or establish a model’s actual knowledge state from calendar dates alone.

Loading and running

from datasets import load_dataset

ds = load_dataset(
    "lfqian/Annulus-Economic-Benchmark",
    "task1_1990",
    split="test",
    token=False,
)

row = ds[0]
prompt = row["input"]  # Plain text: pass this directly to your model.
gold = row["gold"]     # Keep separate for offline scoring.

The prompts are in English. Task 2 now uses four-option value-recall questions built from the existing fact records. The two-column projection contains all 26,460 benchmark questions.

Auxiliary files

The full annotation tables under data/{task}/year_YYYY.parquet retain IDs, metadata and source evidence. The corresponding viewer/{task}/year_YYYY.parquet tables contain only input and gold, in exactly the same row order.

The optional evaluation/{task}/{year}/input.jsonl and gold.jsonl pairs retain IDs for batch runners that join results by ID. Task 2 IDs use {fact_id}/state. The data/facts and model_inputs folders retain the underlying fact records and prompt-only copies. These auxiliary files are not additional tasks, Dataset Viewer configurations, or benchmark questions. Earlier Task 2 versions remain available in the Hugging Face repository's version history.

metadata/release.json records counts, task/year configurations, regional quotas, topic substitutions, validation status and the hashes of numerical audit reports. metadata/alignments.jsonl connects tasks by object and target period. metadata/evaluation_pairs.json lists the optional JSONL pairs. metadata/panels.json describes the sources, and metadata/checksums.json binds the public files. See source notices for attribution and reuse status.

Language

All model-facing instructions, questions, and answer choices are in English. The underlying source values and provenance are preserved. Quoted source excerpts in the auxiliary evidence retain their original language.