Add split data

README.md

@@ -13,6 +13,11 @@ tags:
 pretty_name: SAGE - AoE2 Dataset
 size_categories:
 - 1K<n<10K
+dataset_info:
+  splits:
+    - name: train
+    - name: validation
+    - name: test
 ---
 
 # AoE2 Dataset
@@ -334,7 +339,7 @@ One row per window. Global world state.
 
 ### `events.arrow` / `events.parquet`
 
-One row per discrete game event. Not windowed-events have exact world timestamps.
+One row per discrete game event. Not windowed, events have exact world timestamps.
 
 | Column | Type | Nullable | Description |
 |--------|------|----------|-------------|
@@ -414,7 +419,7 @@ The `visibilityMask` is a bitfield, bit `i` set means player id `i` has (partial
 
 ### `commands.arrow` / `commands.parquet`
 
-One row per issued game command. Not windowed-commands carry exact world timestamps. Uses a wide denormalized schema: most columns are nullable and only populated for the command types that use them.
+One row per issued game command. Not windowed, commands carry exact world timestamps. Uses a wide denormalized schema: most columns are nullable and only populated for the command types that use them.
 
 > **Note:** Command issuer attribution is not always consistent; some commands do not reliably identify their issuing player. Validation via entity ownership is recommended where accuracy matters.
 
@@ -541,6 +546,7 @@ A single armor class or weapon attack entry. Armor and weapon interactions in Ao
 |-------|------|-------------|
 | `type` | Int16 | Armor/weapon class type ID |
 | `value` | Int16 | Armor/attack value for this class |
+---
 
 ### `descriptor.json`
 
@@ -589,7 +595,7 @@ Key properties:
 
 ### `index.json`
 
-A dataset-level index of all processed games. Useful for filtering and selecting games without reading individual game folders.
+A dataset-level index of all processed games. Useful for filtering and selecting games without reading individual game folders. Companion split files `index_train.json`, `index_validation.json`, and `index_test.json` are provided alongside this file.
 
 #### Top-level fields
 
@@ -613,9 +619,12 @@ A dataset-level index of all processed games. Useful for filtering and selecting
 | `durationMs` | number | Match duration in milliseconds |
 | `startTime` | number | Match start time as a Unix timestamp |
 | `startDate` | string | Match start time as ISO 8601 |
+| `split` | string | Dataset split: `train`, `validation`, or `test` |
 | `players` | array | Per-player metadata (see below) |
 | `paths` | object | Relative paths to all associated files (see below) |
 
+> **Splits:** Games are assigned to `train` (80%), `validation` (10%), and `test` (10%) splits via a deterministic hash of the game name. The same game always receives the same split regardless of when it was added to the dataset. The split is provided for reproducibility and convenience, if your use case does not require a held-out set, treat the full `index.json` as a single corpus. The companion files `index_train.json`, `index_validation.json`, and `index_test.json` each follow the same structure as `index.json` and contain only games for that split.
+
 #### Player fields
 
 | Field | Type | Description |