snute/ai_guru-S3

ViBioMIR / MedKB - AI Guru Season 3 Medical Dataset

Complete crawl corpus for AI Guru Season 3: Medical Knowledge Base & QA.

Overview

• Total Pages / Documents: > 4.4 Million pages
• Total Storage: ~55 GB across SQLite databases
• Languages: Vietnamese (vi), Chinese (zh) & English (en - PubMed)
• Primary Source Domains: 97 direct crawled domains + PubMed (240k articles) + Common Crawl & Wayback Machine archives
• Coverage: 100% of reachable missing pages crawled (including Lao Động, Tâm Anh Hospital, VOV, Báo Lạng Sơn, etc.) and PubMed.

Directory Structure

├── db/                   # Direct crawl SQLite databases (97 domains)
│   ├── www.120ask.com.sqlite           (13.7 GB, 918k pages)
│   ├── www.cnkang.com.sqlite           (8.17 GB, 963k pages)
│   ├── www.familydoctor.com.cn.sqlite  (5.66 GB, 445k pages)
│   ├── www.a-hospital.com.sqlite       (2.00 GB)
│   ├── suckhoecongdongonline.vn.sqlite (1.78 GB)
│   ├── thanhnien.vn.sqlite             (1.57 GB)
│   ├── suckhoedoisong.vn.sqlite        (1.32 GB)
│   ├── zhongyibaodian.net.sqlite       (1.04 GB)
│   ├── medlatec.vn.sqlite              (0.93 GB)
│   ├── nhathuoclongchau.com.vn.sqlite  (0.80 GB, 77k pages)
│   ├── laodong.vn.sqlite               (255 MB, 21.7k pages)
│   └── ... (97 domains total)
├── db_cc/                # Common Crawl archives (Long Châu, Lao Động, Tâm Anh, VOV)
├── db_wayback/           # Wayback Machine archives (Zysjonline, Long Châu, etc.)
├── db_alt/               # Alternative mirror archives (Pharmacity)
└── reports/              # Metadata & crawl statistics
    ├── coverage.csv
    ├── missing_pages.csv
    └── missing_pages.md

Database Schema

Each .sqlite file contains a pages table with the following schema:

CREATE TABLE pages (
    id INTEGER PRIMARY KEY,
    url TEXT,
    final_url TEXT,
    status INTEGER,
    kind TEXT,
    ctype TEXT,
    nbytes INTEGER,
    fetched_at REAL,
    attempts INTEGER,
    err TEXT,
    html BLOB
);

How to Load Data in Python

import sqlite3
import zlib

conn = sqlite3.connect("db/laodong.vn.sqlite")
cur = conn.cursor()

for url, html, status in cur.execute("SELECT url, html, status FROM pages LIMIT 5"):
    print(f"URL: {url} (HTTP {status})")
    # html contains raw HTML content (or compressed depending on kind)
    if isinstance(html, bytes):
        try:
            content = html.decode("utf-8")
        except UnicodeDecodeError:
            content = html.decode("latin1", errors="replace")
    else:
        content = str(html)
    print(f"Content length: {len(content)} chars")

conn.close()

Crawled Missing Pages

All target pages from the competition missing_pages.csv were recovered:

• laodong.vn: 21,745 / 21,745 (100%)
• tamanhhospital.vn: 1,381 / 1,381 (100%)
• vov.vn: 2,131 / 2,131 (100%)
• baolangson.vn: 1,981 / 1,981 (100%)