--- language: - en license: other license_name: source-dependent license_link: https://huggingface.co/datasets/lfqian/Annulus-Economic-Benchmark/blob/main/SOURCE_NOTICES.md task_categories: - question-answering - text-generation pretty_name: Annulus Economic Benchmark size_categories: - 10K ## FinDendro recipe replica (2026-10-24) This independently reconstructed version follows the paper recipe. It is **not the original author sample or original model results**. | Reference year | Facts per year | Prompts per year | |---|---:|---:| | 1999–2020 | 48 | 240 | | 2021–2024 | 204 | 1,020 | | 2025 | 204 | 1,020 | | 2026 (January–September) | 153 | 765 | | Total | 2,229 | 11,145 | The sample includes 1,356 macroeconomic facts, 540 stock-price facts and 333 Federal Reserve policy facts. The real observation pool contains 20,130 records. Each fact has five conditions: value, recall, predict, recall_asof, predict_asof. Use the **findendro_recipe** config for all 11,145 browsable prompts, **findendro_facts** for the 2,229 sampled facts, or **findendro_recipe_YYYY** for one reference year. Existing configurations and the repository's default remain available. ```python from datasets import load_dataset prompts = load_dataset("lfqian/Annulus-Economic-Benchmark", "findendro_recipe", split="test") year_2025 = load_dataset("lfqian/Annulus-Economic-Benchmark", "findendro_recipe_2025", split="test") ``` - [Version directory](findendro_recipe_replica/20261024) - [Full archive with raw snapshots and construction code](findendro_recipe_replica/20261024/findendro-replica-20261024.zip) - [Per-year alignment report](findendro_recipe_replica/20261024/metadata/year_alignment.md) - [Complete five-condition example](findendro_recipe_replica/20261024/example_prompts.txt) - [Construction configuration](findendro_recipe_replica/20261024/metadata/config.json) - [Same-month substitute sources](findendro_recipe_replica/20261024/metadata/supplemental_manifest.json) - [Substitute observation checks](findendro_recipe_replica/20261024/metadata/supplement_check.json) For model evaluation use `findendro_recipe_replica/20261024/evaluation/YYYY/input.jsonl`; factual answer mappings are separately stored in `gold.jsonl`. Viewer Parquet files include answer keys for browsing, and should not be passed wholesale to models. Knowledge claims are decoded separately from factual correctness; no ground-truth internal knowledge label is invented. Years follow the fact's reference month. January as-of prompts intentionally specify December 31 of the preceding year. All question/fact/gold joins and yearly partitions passed validation. **October 2025 uses three real same-month substitute observations:** one-, five-, and thirty-year Treasury yields (`GS1`, `GS5`, `GS30`). They occupy the three missing macro sample slots under their own series IDs. The original October `CPIAUCNS`, `CPILFENS` and `UNRATENSA` observations **remain missing**; no CPI or unemployment values are fabricated or imputed. The original missingness is documented by [BLS CPI](https://www.bls.gov/cpi/additional-resources/2025-federal-government-shutdown-impact-cpi.htm) and [BLS CPS](https://www.bls.gov/cps/methods/2025-federal-government-shutdown-impact-cps.htm). This explicit substitution extends the paper recipe and is documented in `metadata/supplemental_manifest.json` and `metadata/supplement_check.json`. The persistence-controlled subset contains 1,335 facts and has an aggregate random baseline of approximately 0.2708. All sources, sampling choices, option generation rules and missingness are documented in the version directory. These are real retrieved historical snapshots; **original first-release vintages have not been independently verified**. Data reuse remains source-dependent.