Add official subset metadata for selected problems

README.md

@@ -43,7 +43,7 @@ Because the full raw dataset is very large, the competition organizers selected
 ## Subsets
 
 - `normal`: 1000 selected problems, chosen programmatically, with 500 ground-truth `TRUE` labels and 500 ground-truth `FALSE` labels
-- `hard`: 200 selected problems, hand-selected by human mathematicians, with 74 ground-truth `TRUE` labels and 126 ground-truth `FALSE` labels
+- `hard`: 200 selected problems co-curated by human mathematicians and AI, with 74 ground-truth `TRUE` labels and 126 ground-truth `FALSE` labels
 - `hard1`: a deduplicated version of the 200-problem `hard` subset, containing 69 unique problems total, with 24 ground-truth `TRUE` labels and 45 ground-truth `FALSE` labels
 - `hard2`: 200 selected problems co-curated by human mathematicians and AI, with 100 ground-truth `TRUE` labels and 100 ground-truth `FALSE` labels
 
@@ -60,6 +60,23 @@ Each record has the following fields:
 - `equation2`: Equation 2
 - `answer`: whether Equation 1 implies Equation 2 over all magmas
 
+## Subset Metadata
+
+Subset-level metadata for the official released subsets is stored under `metadata/`.
+
+Each metadata file contains:
+
+- `subset_name`: subset identifier
+- `source`: release source, such as `official`
+- `curators`: list of named curators, if provided
+- `difficulty`: subset difficulty label
+- `problem_count`: total number of problems in the subset
+- `true_count`: number of ground-truth `TRUE` labels
+- `false_count`: number of ground-truth `FALSE` labels
+- `selection_method`: high-level selection method such as `programmatic`, `human`, or `human_ai`
+- `derived_from`: parent subset name if the subset is derived from another subset
+- `notes`: optional explanatory notes
+
 Records in `hard1` use `hard1_*` IDs with `difficulty: hard`, and records in `hard2` use `hard2_*` IDs with `difficulty: hard`.
 
 ## Usage
@@ -98,3 +115,7 @@ hard2 = load_dataset(
 - `data/hard.jsonl`
 - `data/hard1.jsonl`
 - `data/hard2.jsonl`
+- `metadata/normal.json`
+- `metadata/hard.json`
+- `metadata/hard1.json`
+- `metadata/hard2.json`

metadata/hard.json

@@ -0,0 +1,12 @@
+{
+  "subset_name": "hard",
+  "source": "official",
+  "curators": [],
+  "difficulty": "hard",
+  "problem_count": 200,
+  "true_count": 74,
+  "false_count": 126,
+  "selection_method": "human_ai",
+  "derived_from": null,
+  "notes": ""
+}

metadata/hard1.json

@@ -0,0 +1,12 @@
+{
+  "subset_name": "hard1",
+  "source": "official",
+  "curators": [],
+  "difficulty": "hard",
+  "problem_count": 69,
+  "true_count": 24,
+  "false_count": 45,
+  "selection_method": "human_ai",
+  "derived_from": "hard",
+  "notes": "Deduplicated version of the hard subset; contains no duplicate problems."
+}

metadata/hard2.json

@@ -0,0 +1,12 @@
+{
+  "subset_name": "hard2",
+  "source": "official",
+  "curators": [],
+  "difficulty": "hard",
+  "problem_count": 200,
+  "true_count": 100,
+  "false_count": 100,
+  "selection_method": "human_ai",
+  "derived_from": null,
+  "notes": ""
+}

metadata/normal.json

@@ -0,0 +1,12 @@
+{
+  "subset_name": "normal",
+  "source": "official",
+  "curators": [],
+  "difficulty": "normal",
+  "problem_count": 1000,
+  "true_count": 500,
+  "false_count": 500,
+  "selection_method": "programmatic",
+  "derived_from": null,
+  "notes": ""
+}