Show exactly input and gold columns for every task and year

README.md

@@ -22,161 +22,161 @@ configs:
   default: true
   data_files:
   - split: year_1990
-    path: data/task1/year_1990.parquet
+    path: viewer/task1/year_1990.parquet
   - split: year_1991
-    path: data/task1/year_1991.parquet
+    path: viewer/task1/year_1991.parquet
   - split: year_1992
-    path: data/task1/year_1992.parquet
+    path: viewer/task1/year_1992.parquet
   - split: year_1993
-    path: data/task1/year_1993.parquet
+    path: viewer/task1/year_1993.parquet
   - split: year_1994
-    path: data/task1/year_1994.parquet
+    path: viewer/task1/year_1994.parquet
   - split: year_1995
-    path: data/task1/year_1995.parquet
+    path: viewer/task1/year_1995.parquet
   - split: year_1996
-    path: data/task1/year_1996.parquet
+    path: viewer/task1/year_1996.parquet
   - split: year_1997
-    path: data/task1/year_1997.parquet
+    path: viewer/task1/year_1997.parquet
   - split: year_1998
-    path: data/task1/year_1998.parquet
+    path: viewer/task1/year_1998.parquet
   - split: year_1999
-    path: data/task1/year_1999.parquet
+    path: viewer/task1/year_1999.parquet
   - split: year_2000
-    path: data/task1/year_2000.parquet
+    path: viewer/task1/year_2000.parquet
   - split: year_2001
-    path: data/task1/year_2001.parquet
+    path: viewer/task1/year_2001.parquet
   - split: year_2002
-    path: data/task1/year_2002.parquet
+    path: viewer/task1/year_2002.parquet
   - split: year_2003
-    path: data/task1/year_2003.parquet
+    path: viewer/task1/year_2003.parquet
   - split: year_2004
-    path: data/task1/year_2004.parquet
+    path: viewer/task1/year_2004.parquet
   - split: year_2005
-    path: data/task1/year_2005.parquet
+    path: viewer/task1/year_2005.parquet
   - split: year_2006
-    path: data/task1/year_2006.parquet
+    path: viewer/task1/year_2006.parquet
   - split: year_2007
-    path: data/task1/year_2007.parquet
+    path: viewer/task1/year_2007.parquet
   - split: year_2008
-    path: data/task1/year_2008.parquet
+    path: viewer/task1/year_2008.parquet
   - split: year_2009
-    path: data/task1/year_2009.parquet
+    path: viewer/task1/year_2009.parquet
   - split: year_2010
-    path: data/task1/year_2010.parquet
+    path: viewer/task1/year_2010.parquet
   - split: year_2011
-    path: data/task1/year_2011.parquet
+    path: viewer/task1/year_2011.parquet
   - split: year_2012
-    path: data/task1/year_2012.parquet
+    path: viewer/task1/year_2012.parquet
   - split: year_2013
-    path: data/task1/year_2013.parquet
+    path: viewer/task1/year_2013.parquet
   - split: year_2014
-    path: data/task1/year_2014.parquet
+    path: viewer/task1/year_2014.parquet
   - split: year_2015
-    path: data/task1/year_2015.parquet
+    path: viewer/task1/year_2015.parquet
   - split: year_2016
-    path: data/task1/year_2016.parquet
+    path: viewer/task1/year_2016.parquet
   - split: year_2017
-    path: data/task1/year_2017.parquet
+    path: viewer/task1/year_2017.parquet
   - split: year_2018
-    path: data/task1/year_2018.parquet
+    path: viewer/task1/year_2018.parquet
   - split: year_2019
-    path: data/task1/year_2019.parquet
+    path: viewer/task1/year_2019.parquet
   - split: year_2020
-    path: data/task1/year_2020.parquet
+    path: viewer/task1/year_2020.parquet
   - split: year_2021
-    path: data/task1/year_2021.parquet
+    path: viewer/task1/year_2021.parquet
   - split: year_2022
-    path: data/task1/year_2022.parquet
+    path: viewer/task1/year_2022.parquet
   - split: year_2023
-    path: data/task1/year_2023.parquet
+    path: viewer/task1/year_2023.parquet
   - split: year_2024
-    path: data/task1/year_2024.parquet
+    path: viewer/task1/year_2024.parquet
   - split: year_2025
-    path: data/task1/year_2025.parquet
+    path: viewer/task1/year_2025.parquet
   - split: year_2026
-    path: data/task1/year_2026.parquet
+    path: viewer/task1/year_2026.parquet
 - config_name: task2
   data_files:
   - split: year_1990
-    path: data/task2/year_1990.parquet
+    path: viewer/task2/year_1990.parquet
   - split: year_1991
-    path: data/task2/year_1991.parquet
+    path: viewer/task2/year_1991.parquet
   - split: year_1992
-    path: data/task2/year_1992.parquet
+    path: viewer/task2/year_1992.parquet
   - split: year_1993
-    path: data/task2/year_1993.parquet
+    path: viewer/task2/year_1993.parquet
   - split: year_1994
-    path: data/task2/year_1994.parquet
+    path: viewer/task2/year_1994.parquet
   - split: year_1995
-    path: data/task2/year_1995.parquet
+    path: viewer/task2/year_1995.parquet
   - split: year_1996
-    path: data/task2/year_1996.parquet
+    path: viewer/task2/year_1996.parquet
   - split: year_1997
-    path: data/task2/year_1997.parquet
+    path: viewer/task2/year_1997.parquet
   - split: year_1998
-    path: data/task2/year_1998.parquet
+    path: viewer/task2/year_1998.parquet
   - split: year_1999
-    path: data/task2/year_1999.parquet
+    path: viewer/task2/year_1999.parquet
   - split: year_2000
-    path: data/task2/year_2000.parquet
+    path: viewer/task2/year_2000.parquet
   - split: year_2001
-    path: data/task2/year_2001.parquet
+    path: viewer/task2/year_2001.parquet
   - split: year_2002
-    path: data/task2/year_2002.parquet
+    path: viewer/task2/year_2002.parquet
   - split: year_2003
-    path: data/task2/year_2003.parquet
+    path: viewer/task2/year_2003.parquet
   - split: year_2004
-    path: data/task2/year_2004.parquet
+    path: viewer/task2/year_2004.parquet
   - split: year_2005
-    path: data/task2/year_2005.parquet
+    path: viewer/task2/year_2005.parquet
   - split: year_2006
-    path: data/task2/year_2006.parquet
+    path: viewer/task2/year_2006.parquet
   - split: year_2007
-    path: data/task2/year_2007.parquet
+    path: viewer/task2/year_2007.parquet
   - split: year_2008
-    path: data/task2/year_2008.parquet
+    path: viewer/task2/year_2008.parquet
   - split: year_2009
-    path: data/task2/year_2009.parquet
+    path: viewer/task2/year_2009.parquet
   - split: year_2010
-    path: data/task2/year_2010.parquet
+    path: viewer/task2/year_2010.parquet
   - split: year_2011
-    path: data/task2/year_2011.parquet
+    path: viewer/task2/year_2011.parquet
   - split: year_2012
-    path: data/task2/year_2012.parquet
+    path: viewer/task2/year_2012.parquet
   - split: year_2013
-    path: data/task2/year_2013.parquet
+    path: viewer/task2/year_2013.parquet
   - split: year_2014
-    path: data/task2/year_2014.parquet
+    path: viewer/task2/year_2014.parquet
   - split: year_2015
-    path: data/task2/year_2015.parquet
+    path: viewer/task2/year_2015.parquet
   - split: year_2016
-    path: data/task2/year_2016.parquet
+    path: viewer/task2/year_2016.parquet
   - split: year_2017
-    path: data/task2/year_2017.parquet
+    path: viewer/task2/year_2017.parquet
   - split: year_2018
-    path: data/task2/year_2018.parquet
+    path: viewer/task2/year_2018.parquet
   - split: year_2019
-    path: data/task2/year_2019.parquet
+    path: viewer/task2/year_2019.parquet
   - split: year_2020
-    path: data/task2/year_2020.parquet
+    path: viewer/task2/year_2020.parquet
   - split: year_2021
-    path: data/task2/year_2021.parquet
+    path: viewer/task2/year_2021.parquet
   - split: year_2022
-    path: data/task2/year_2022.parquet
+    path: viewer/task2/year_2022.parquet
   - split: year_2023
-    path: data/task2/year_2023.parquet
+    path: viewer/task2/year_2023.parquet
   - split: year_2024
-    path: data/task2/year_2024.parquet
+    path: viewer/task2/year_2024.parquet
   - split: year_2025
-    path: data/task2/year_2025.parquet
+    path: viewer/task2/year_2025.parquet
   - split: year_2026
-    path: data/task2/year_2026.parquet
+    path: viewer/task2/year_2026.parquet
 - config_name: task3
   data_files:
   - split: year_2025
-    path: data/task3/year_2025.parquet
+    path: viewer/task3/year_2025.parquet
   - split: year_2026
-    path: data/task3/year_2026.parquet
+    path: viewer/task3/year_2026.parquet
 ---
 
 # Annulus Economic Benchmark
@@ -206,11 +206,18 @@ The data are constructed from frozen statistical-table downloads from the Bank f
 
 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.
 
-## Files and use
+## Two-column evaluation table
 
-For model runs, use the separate `evaluation/{task}/{year}/input.jsonl` and `gold.jsonl` files described below.
+Select **Subset** (`task1`, `task2`, or `task3`), then **Split** (`year_YYYY`). Every displayed row has exactly two top-level columns:
 
-The Dataset Viewer provides three configurations: `task1`, `task2`, and `task3`. Each is partitioned into target-year splits such as `year_1990` and `year_2025`. Question records include `messages` for model inference and separate scoring metadata. Gold answers, conditional labels, and evidence are supplied for offline evaluation; they must not be included in model input. Use only `messages` when constructing an inference request, and join predictions to scoring records by their IDs.
+| 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 correct trajectory option letter. Task 3's `gold` is the observed outcome's option letter; its input still requests a probability distribution and a reason. The 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 span `year_1990` through `year_2026`. Task 3 has `year_2025` and `year_2026`. The month remains written in each monthly question. Year splits organize evaluation items by target year, not training/validation partitions.
 
 ## Validation and limitations
 
@@ -220,59 +227,29 @@ The source snapshots may contain revisions and are not historical real-time vint
 
 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.
 
-## Model input and gold files
-
-Use the separate JSONL pair for each task and target year to run inference and score predictions:
-
-```text
-evaluation/task1/1990/input.jsonl
-evaluation/task1/1990/gold.jsonl
-evaluation/task2/1990/input.jsonl
-evaluation/task2/1990/gold.jsonl
-evaluation/task3/2026/input.jsonl
-evaluation/task3/2026/gold.jsonl
-```
-
-Every `input.jsonl` row has only `id` and `messages`. Send only `messages` to the model; keep `id` in the client so its output can be joined to gold. The corresponding `gold.jsonl` has the same IDs in the same order, with offline answer keys and scoring metadata. Never include a gold row in a model request. The 76 task/year partitions contain 26,460 inputs and 26,460 corresponding gold records; these are separate views of the existing questions, not new questions. The complete file list and counts are in `metadata/evaluation_pairs.json`.
-
-Task 1 gold retains `mode`, the option-letter-to-behavior mapping, and `gold_by_boundary` with `inside` and `outside` option letters. Choose the condition using your evaluation protocol; the files do not infer an individual model's knowledge from a calendar date. Task 2 uses `gold_label` for the correct trajectory option. Task 3 uses `gold_label` for the observed outcome and retains the original `choices`, `status`, `target_month` and `topic_id` needed for probability scoring.
-
-Task 1 and Task 2 model responses are option letters. Task 3 responses contain `probabilities` for the actual option letters and a nonempty `reason`; preserve these probabilities rather than keeping only their maximum. The client should attach the input's `id` when saving responses. There is no single reference reasoning text in gold.
-
-Download one input/gold pair:
+## Loading and running
 
 ```python
-import json
-from huggingface_hub import hf_hub_download
-
-repo_id = "lfqian/Annulus-Economic-Benchmark"
-prefix = "evaluation/task1/1990"
-input_path = hf_hub_download(repo_id, f"{prefix}/input.jsonl", repo_type="dataset")
-gold_path = hf_hub_download(repo_id, f"{prefix}/gold.jsonl", repo_type="dataset")
+from datasets import load_dataset
 
-with open(input_path) as f:
-    inputs = [json.loads(line) for line in f]
-with open(gold_path) as f:
-    gold_by_id = {row["id"]: row for row in map(json.loads, f)}
+ds = load_dataset(
+    "lfqian/Annulus-Economic-Benchmark",
+    "task1",
+    split="year_1990",
+    token=False,
+)
 
-messages = inputs[0]["messages"]  # Send this to your model.
-gold = gold_by_id[inputs[0]["id"]]  # Read this only for offline scoring.
+row = ds[0]
+prompt = row["input"]  # Plain text: pass this directly to your model.
+gold = row["gold"]     # Keep separate for offline scoring.
 ```
 
-## Dataset Viewer loading
-
-```python
-from datasets import load_dataset
-
-questions_1990 = load_dataset("lfqian/Annulus-Economic-Benchmark", "task1", split="year_1990")
-forecasts_2026 = load_dataset("lfqian/Annulus-Economic-Benchmark", "task3", split="year_2026")
-september_2026 = forecasts_2026.filter(lambda x: x["period"] == "2026-09")
-```
+The input text, option order and answers are unchanged from the source-derived release. The two-column projection contains all 26,460 benchmark questions.
 
-The three task configurations contain the question text, labelled `choices`, temporal and geographical metadata, `messages` for inference, and offline scoring fields. The `model_inputs/task1`, `model_inputs/task2`, and `model_inputs/task3` folders contain optional prompt-only Parquet files with just `id` and `messages`, organized by the same target years. The `data/facts` folder contains the underlying numerical observations and numerical-answer choices. These auxiliary files are available through the repository file browser; they are not additional Dataset Viewer configurations, tasks, or benchmark questions.
+## Auxiliary files
 
-In Task 1, `gold_inside_label` and `gold_outside_label` are option letters under the two boundary conditions. In Tasks 2 and 3, `gold_label` is the observed answer's option letter. Nested source evidence is serialized as `evidence_json` to preserve heterogeneous source metadata. Task 1's `option_metadata_json` identifies each option's response behaviour.
+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 first Dataset Viewer selector, **Subset**, selects the task/configuration. The second, **Split**, selects a target year named `year_YYYY`. Task 1 and Task 2 span `year_1990` through `year_2026`; Task 3 has only `year_2025` and `year_2026`. Auxiliary facts and prompt-only files use corresponding year partitions. Within 2025 and 2026, filter `period` (`YYYY-MM`) for a particular month. These splits organize evaluation items by target year; they are not training/validation partitions. Task 1 recall/predict questions remain distinguished by `mode` within each year. `alignment_key` connects records for the same object and target period; `metadata/alignments.jsonl` lists the joins explicitly.
+The optional `evaluation/{task}/{year}/input.jsonl` and `gold.jsonl` pairs retain IDs for batch runners that join results by ID. 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.
 
-`metadata/release.json` records exact counts, per-period distributions, regional quotas, topic substitutions, and the hashes of the numerical audit reports used for this release. `metadata/panels.json` describes the source panels, and `metadata/checksums.json` binds the public files. See [source notices](SOURCE_NOTICES.md) for attribution and reuse status.
+`metadata/release.json` records counts, year splits, 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](SOURCE_NOTICES.md) for attribution and reuse status.