DATA_DICTIONARY.md

FRED-SemBench V1 Candidate Data Dictionary

concepts.json

• concept_id: frozen benchmark identifier from FRED-C001 to FRED-C050.
• name: human-readable indicator concept.
• series_id: canonical FRED series used by the benchmark.
• frequency: reported observation frequency.
• official_source: agency responsible for the underlying statistic.
• unit / display_unit: normalized and human-readable units.
• precision: decimal quantization used for gold answers.
• change_kind: deterministic rule used to construct the change fact.
• alternative_series_id: plausible confounder used in disambiguation cases.

facts.jsonl

Each concept has a level fact and a change fact.

• fact_id: stable fact identifier.
• benchmark_concept_id: foreign key to concepts.json.
• series_ids: required provider series.
• observation_period / observation_dates: human and ISO representations.
• as_of: benchmark snapshot cutoff.
• vintage_mode: currently frozen_current_vintage.
• raw_observations: values used by the deterministic derivation.
• derivation: transform name and formula description.
• answer: numeric gold, unit, and precision.
• source: official source, provider, and series URL.
• snapshot: request URL, local file, snapshot ID, and SHA-256.
• verification: automatic recomputation status.

questions.jsonl

• case_id: stable ID from fred_sem_0001 to fred_sem_0200.
• primary_concept_id: concept used for allocation and stratification.
• task_type: one of the six frozen task types.
• fact_ids: evidence records; empty for clarification cases.
• gold.answer_mode: numeric or clarification.
• gold: concept, series, transform, period, answer, provenance, and required response elements.
• confounders: alternative series and the expected common error.
• annotation: review status and review notes.

Review files

• allocation.csv: one row per concept and one binary column per task type.
• questions_review.csv: flattened review view; reviewers complete reviewer_decision and reviewer_notes.
• QA_REPORT.md: automated results and unresolved release gates.

Snapshot semantics

The CSV files reproduce the provider responses downloaded on 2026-07-29. They are frozen current-vintage evidence, not historical ALFRED vintages. A future historical-vintage extension must use a distinct benchmark version.