Update dataset card

README.md

@@ -10,34 +10,12 @@ configs:
 - config_name: default
   data_files:
   - split: train
-    path: data/train-*
+    path: data/train-*.parquet
 tags:
 - simple-wikipedia
 - wikipedia
 - markdown
 - sqlite
-dataset_info:
-  features:
-  - name: page_id
-    dtype: int64
-  - name: title
-    dtype: string
-  - name: content
-    dtype: string
-  - name: content_no_link
-    dtype: string
-  - name: importance
-    dtype: string
-  - name: truncated
-    dtype: bool
-  - name: error
-    dtype: bool
-  splits:
-  - name: train
-    num_bytes: 797840092
-    num_examples: 273522
-  download_size: 432389310
-  dataset_size: 797840092
 ---
 
 # Simple English Wikipedia (Markdown)
@@ -52,6 +30,23 @@ Recurring weekly snapshot of Simple English Wikipedia (https://simple.wikipedia.
 
 ## Dataset Structure
 
+- Hugging Face split: `train`
+- SQLite mirror: `simplewiki.sqlite` (contains `pages` and `infobox` tables)
+
+## Hugging Face Dataset Columns (train split)
+
+- `page_id` (int64): Unique page identifier from Wikimedia dump.
+- `title` (string): Article title.
+- `content` (string): Article body converted to markdown with internal and external links preserved.
+- `content_no_link` (string): Same content with markdown links stripped to plain text.
+- `importance` (string): Importance for a household smart speaker assistant (`low`, `medium`, `high`, or `unknown` when not categorized).
+- `truncated` (bool): `true` when the source article exceeded 40,000 characters; only the first two paragraphs are parsed and stored in this case; otherwise `false`.
+- `error` (bool): `true` when the article could not be parsed (content fields are empty in this case); otherwise `false`.
+
+## SQLite Database
+
+### `pages` table
+
 Columns:
 - `page_id` (int64): Unique page identifier from Wikimedia dump.
 - `title` (string): Article title.
@@ -61,13 +56,38 @@ Columns:
 - `truncated` (bool): `true` when the source article exceeded 40,000 characters; only the first two paragraphs are parsed and stored in this case; otherwise `false`.
 - `error` (bool): `true` when the article could not be parsed (content fields are empty in this case); otherwise `false`.
 
+Example:
+
+```bash
+sqlite3 simplewiki.sqlite "SELECT page_id, title, substr(content,1,200) || '...' AS preview FROM pages ORDER BY page_id LIMIT 5;"
+```
+
+### `infobox` table
+
+Columns:
+- `page_id` (int64): Page ID of the article the infobox belongs to.
+- `article_title` (string): Title of the article.
+- `infobox_type` (string): Template type (e.g., `country`, `person`, `airport`).
+- `label` (string): Raw parameter name from the infobox template.
+- `value` (string): Cleaned text value with wiki links flattened to plain text.
+
+Examples:
+
+```bash
+# All fields for Australia
+sqlite3 simplewiki.sqlite "SELECT label, value FROM infobox WHERE page_id = 27 ORDER BY label LIMIT 10;"
+
+# Join back to article text
+sqlite3 simplewiki.sqlite "SELECT p.title, i.label, i.value FROM infobox i JOIN pages p ON p.page_id = i.page_id WHERE i.infobox_type = 'country' AND i.label = 'capital' LIMIT 5;"
+```
+
 ## Processing
 
 - Downloaded `pages-articles` XML dump and verified SHA-1.
 - Kept only namespace 0 articles, skipped redirects, and dropped titles beginning with “List of”.
 - Stripped templates/ref/gallery blocks and file/category links; converted headings, lists, tables, and internal/external links to Markdown with page IDs.
-- Stored a SQLite mirror (`pages` table) alongside the Hugging Face dataset.
-- Markdown links point to the target page's numeric ID for fast lookup without a title-to-ID join.
+- Stored a SQLite mirror with both `pages` and `infobox` tables; infobox templates are flattened into rows with cleaned text values.
+- Markdown links point to the target page's numeric ID for fast lookup without a title-to-ID join; infobox values have links stripped to plain text for ease of use.
 
 ## Usage
 
@@ -80,10 +100,13 @@ ds = load_dataset("juno-labs/simple_wikipedia", split="train")
 print(ds[0])
 ```
 
-SQLite usage (`simplewiki.sqlite` mirrors the same columns):
+SQLite usage (`simplewiki.sqlite` mirrors the same columns plus `infobox`):
 
 ```bash
 sqlite3 simplewiki.sqlite "SELECT page_id, title, substr(content,1,200) || '...' FROM pages LIMIT 5;"
+
+# Infobox lookup example (Australia):
+sqlite3 simplewiki.sqlite "SELECT label, value FROM infobox WHERE page_id = 27 ORDER BY label LIMIT 5;"
 ```
 
 You can also mount it in code:
@@ -99,7 +122,14 @@ for row in cur.execute("SELECT title, content FROM pages WHERE page_id = ?", (75
 
 ## Categorization
 
-- Model: openai/gpt-5-mini
+Importance labels indicate how useful an article is for day-to-day offline household smart-speaker queries; `unknown` is used when labeling is disabled or fails.
+
+- Model: openai/gpt-oss-120b
+- distribution:
+  - low: 67.78%
+  - medium: 29.85%
+  - high: 2.35%
+  - unknown: 0.02%
 
 Prompt template:
 ```