Datasets:

SafeVixAI
/

SafeVixAI-Dataset-Hub

license: apache-2.0
task_categories:
  - object-detection
  - question-answering
language:
  - en
  - hi
tags:
  - road-safety
  - india
  - emergency-services
  - traffic-law
  - geospatial
  - rag
pretty_name: SafeVixAI Dataset Hub
size_categories:
  - 1B<n<10B

SafeVixAI Dataset Hub 🛡️

The Intelligence Layer for the SafeVixAI platform — IIT Madras Road Safety Hackathon 2026

This repository hosts all datasets, pre-trained models, notebooks, and reproducible data acquisition scripts that power the SafeVixAI application. It is designed to be cloned directly into Google Colab or any research environment.

Main Application Repo: SafeVixAI/SafeVixAI

⚡ Quickstart (Google Colab)

# Clone the entire intelligence layer
!git clone https://huggingface.co/datasets/SafeVixAI/SafeVixAI-Dataset-Hub /content/data

# Symlink into the app structure
import os
os.makedirs("/content/SafeVixAI/chatbot_service", exist_ok=True)
!ln -sfn /content/data/data/chatbot_service/data /content/SafeVixAI/chatbot_service/data

📦 Repository Structure

SafeVixAI-Dataset-Hub/
├── data/                         ← 4.2 GB of raw intelligence data (11,008 files)
│   ├── backend/
│   │   ├── data/                 ← Civic intel (OSM data, toll plazas, ward boundaries)
│   │   └── datasets/             ← Raw database assets (blackspots, violations CSVs)
│   ├── chatbot_service/data/     ← Chatbot reference directories & built vector stores
│   └── frontend/offline-data/    ← Regional translation matrices & offline PWA bundles
│
└── scripts/                      ← Complete data acquisition & processing pipelines
    ├── backend/                  ← Core backend and database migration scripts
    ├── chatbot_service/          ← Chatbot agent and QA validation scripts
    └── scripts/                  ← Legacy scrapers, downloaders, and seeders

🗂️ Data Contents

Category	Files	Size	Source
Emergency Services GIS	`hospitals.csv`, `police_stations.csv`, `fire_stations.csv`, `ambulance.csv`	~150 MB	OpenStreetMap Overpass API
Legal Documents	`motor_vehicles_act_1988.pdf`, `mv_amendment_act_2019.pdf`	~30 MB	MoRTH / Indian Kanoon
Accident Statistics	`kaggle_india_accidents.csv`, `morth_2022/*.csv`	~250 MB	MoRTH / Kaggle
Road Infrastructure	`pmgsy_roads.geojson`, `toll_plazas.csv`	~900 MB	PMGSY GeoSadak / NHAI
Hospital Directory	`hospital_directory.csv`, `nin_facilities.csv`	~1.2 GB	NHP / NIN
Traffic Violations	`violations_seed.csv`, `state_overrides.csv`	~5 MB	MVA 2019
PWA Translations	`translations/*.json` (11 regional Indian languages)	~6.4 MB	Automated DeepL/Google Sync
Road Damage Model	`road_damage_2025/`	~800 MB	ONNX + Training Data
QA Pairs	`qa_pairs/`	~50 MB	Custom RAG Training

🔬 Scripts — Reproducible Data Pipeline

The scripts/ folder mirrors the main application's data acquisition pipeline. All scripts here are pure Python — they run without any database or backend stack.

Structure

scripts/
├── scripts/data/             ← from SafeVixAI/scripts/data/
│   ├── _overpass_utils.py    ← Core GIS utility (basic version)
│   ├── fetch_hospitals.py    ← Hospital data from OpenStreetMap
│   ├── fetch_police.py       ← Police station data
│   ├── fetch_ambulance.py    ← Ambulance service data
│   ├── fetch_blood_banks.py  ← Blood bank data
│   ├── fetch_fire.py         ← Fire station data
│   ├── download_legal_pdfs.py ← MV Act PDF downloader
│   ├── extract_morth2022_tables.py ← PDF → CSV extractor
│   ├── seed_blackspots.py    ← Accident blackspot normalizer
│   ├── bootstrap_local_data.py ← Master data orchestrator
│   ├── verify_data.py        ← Data integrity checker
│   ├── inspect_zips.py       ← Dataset zip validator
│   ├── audit_env.py          ← Environment configuration checker
│   ├── check_all_scripts.py  ← Script syntax validator
│   └── setup_kaggle.ps1      ← Kaggle API auth setup
│
├── backend/data/             ← from SafeVixAI/backend/scripts/data/
│   ├── seed_violations.py    ← Traffic fine normalizer (MVA 2019)
│   ├── prepare_road_sources.py ← CSV/GeoJSON → LineString converter
│   ├── sample_pmgsy.py       ← PMGSY 867K roads sampler
│   ├── road_sources.example.json ← Manifest template
│   └── road_sources.json     ← Active road source manifest
│
└── chatbot_service/data/     ← from SafeVixAI/chatbot_service/scripts/data/
    ├── _overpass_utils.py    ← Pro GIS utility (with retries + backoff)
    ├── fetch_hospitals.py    ← Pro hospital fetcher
    ├── fetch_police.py       ← Pro police fetcher
    ├── fetch_ambulance.py    ← Pro ambulance fetcher
    ├── fetch_blood_banks.py  ← Pro blood bank fetcher
    └── fetch_fire.py         ← Pro fire station fetcher

Running the Scripts

# 1. Setup Kaggle auth (one-time)
pwsh scripts/scripts/data/setup_kaggle.ps1

# 2. Fetch all emergency service GIS data (Pro versions recommended)
python scripts/chatbot_service/data/fetch_hospitals.py
python scripts/chatbot_service/data/fetch_police.py
python scripts/chatbot_service/data/fetch_ambulance.py
python scripts/chatbot_service/data/fetch_blood_banks.py
python scripts/chatbot_service/data/fetch_fire.py

# 3. Download legal documents
python scripts/scripts/data/download_legal_pdfs.py

# 4. Extract MoRTH accident tables
python scripts/scripts/data/extract_morth2022_tables.py

# 5. Verify everything
python scripts/scripts/data/verify_data.py

Note: chatbot_service/data/ versions of the fetchers are the Pro versions — they include retry logic, exponential backoff, and more precise Indian GIS selectors.

📓 Research Notebooks — Open in Colab

We advise running all notebooks through Google Colab for the easiest setup. Colab gives you a free T4 GPU with 16 GB of VRAM. All notebooks were built and tested on Colab, so it is the most stable platform. Any other cloud provider should work too.

#	Notebook	What It Produces
1	YOLOv8 Pothole Detector Training	ONNX road damage model
2	ChromaDB RAG Vectorstore Build	ChromaDB index for legal RAG
3	Accident EDA & Hotspot Generator	Blackspot seed CSV + heatmap
4	Roads Data Processing	Sampled PMGSY GeoJSON
5	Risk Model ONNX Training	Risk scoring ONNX model

Recommended run order: 1 → 2 → 3 → 4 → 5

Each notebook auto-clones this Hub at the start — no manual data setup needed.

📜 License

Apache 2.0 — See LICENSE

Data sourced from OpenStreetMap (ODbL), MoRTH (Government of India Open Data), Kaggle, PMGSY GeoSadak, and National Health Portal.