NeroSeungSan/synthengine-cot-edge-case-v1

SynthEngine CoT Edge Case Dataset v1.0

Premium synthetic Chain-of-Thought reasoning data for autonomous driving, robotics, and embodied AI edge cases.

🔗 Full dataset (1000 records) available on Gumroad This HuggingFace repo contains a free sample (10 records) under CC BY-NC-SA 4.0.

🎯 Why This Dataset?

In 2025, NVIDIA Alpamayo-R1 proved that Chain-of-Causation reasoning improves autonomous driving planning accuracy by +12% and reduces close encounters by -35%. In 2024, UC Berkeley's Embodied CoT showed that reasoning augmentation boosts robot policy success rates by +28%.

The industry is shifting from perception-only data to reasoning data. SynthEngine is purpose-built for this new paradigm.

📊 Dataset Statistics

Metric	Value
Total Records	1000
Format	JSONL (one JSON object per line)
Avg CoT Steps	5.9 steps/record
Categories	7 scenario domains
Severity Levels	4 (low/medium/high/critical)
Size	~1.5 MB
Generation Model	DeepSeek Chat (temperature 0.85)
Version	1.0.0

🏷️ Scenario Categories

Category	Records	Description
adverse_weather	204 (20.4%)	惡劣天氣行車 — Heavy rain, fog, night driving, hydroplaning
pedestrian_dart_out	182 (18.2%)	行人鬼探頭 — Pedestrian emergence from occluded spots
fragile_grasp	144 (14.4%)	機械臂精密抓取 — Force-controlled fragile object grasping
motorcycle_cut_in	140 (14.0%)	電單車切入 — Sudden cut-in from adjacent traffic
child_heat_safety	139 (13.9%)	兒童高溫安全 — Child near high-temperature hazard reasoning
pet_avoidance	99 (9.9%)	寵物避讓 — Unexpected pet crossing robot path
falling_cargo	88 (8.8%)	貨物掉落 — Large object falling from leading vehicle
generic_edge_case	4 (0.4%)	Unclassified edge cases

📋 Schema

Each record contains the following fields:

Field	Type	Description
scenario_id	string	Unique ID (SynthEngine-XXXX)
scenario_type	string	Normalised category label
scenario_type_zh	string	Chinese category label
scenario_type_original	string	Original raw label (before normalisation)
environment_description	string	Detailed scene description (67-709 chars)
chain_of_thought	string[]	4-8 step reasoning chain
force_control_action	string	Executable control directive
severity	string	low, medium, high, or critical
severity_numeric	int	Numeric severity (1-4)
version	string	Dataset version
source_model	string	Generation model
source_temperature	float	Generation temperature
product	string	Dataset name

🔬 Sample Record

{
  "scenario_id": "SynthEngine-0001",
  "scenario_type": "adverse_weather",
  "scenario_type_zh": "惡劣天氣行車",
  "environment_description": "Torrential rain at night on a highway with visibility under 20 meters...",
  "chain_of_thought": [
    "Assess visibility: less than 20 meters, cannot see lane markings or obstacles ahead.",
    "Check road surface: standing water and hydroplaning likely; reduce speed to below 20 km/h.",
    "Evaluate wind: strong gusts require steering countersteer to maintain lane position.",
    "Decide to turn on hazard lights and fog lights to increase visibility to others.",
    "Determine that stopping is unsafe due to risk of rear-end collision; continue at slow pace toward nearest exit.",
    "If exit not reachable, pull over as far right as possible and stop with wheels straight, engine off, foot off brake to avoid brake light confusion."
  ],
  "force_control_action": "Gradual deceleration to 15 km/h, engage hazard lights, maintain firm but gentle steering input...",
  "severity": "critical",
  "severity_numeric": 4
}

🛠️ Use Cases

• Fine-tuning VLAs (Vision-Language-Action models) for autonomous driving planning
• Training embodied reasoning for robot policies (OpenVLA, RT-2, Octo)
• Edge case augmentation for AV simulation testing
• Safety reasoning evaluation for autonomous systems
• Chain-of-Thought distillation for small reasoning models

🎓 Related Research

• NVIDIA Alpamayo-R1: Chain of Causation for Autonomous Driving — +12% planning accuracy
• Embodied Chain-of-Thought Reasoning (UC Berkeley/Stanford) — +28% robot success rate

📄 License

• Sample (this repo): CC BY-NC-SA 4.0 (10 free records)
• Full dataset: Commercial License — available on Gumroad

📬 Contact

• Website: synthengine-data.com
• For enterprise licensing / custom generation: contact via website