Ashith18/sentinel-disasternet-grpo

Uploaded model

🚨 SENTINEL: Teaching AI to Coordinate Earthquake Disaster Response with GRPO

🚀 GRPO training improved reward by +27% in 300 steps. The model discovered hospital protection (89% vs 31%) without being programmed.

🌍 The Problem

In the 2023 Turkey–Syria earthquake, over 50,000 people lost their lives.

Many deaths were preventable — not due to lack of resources, but due to poor coordination in the first 72 hours.

🧠 What is DisasterNET?

DisasterNET is a reinforcement learning environment simulating real-world disaster response.

• —10 city zones (hospital, schools, residential areas)
• —12 sequential decisions across 72 hours
• —Cascade failure (hospital fails if not protected)
• —Fog of war (only 40% data visible)
• —Based on real earthquake data

⚙️ Training Setup

Model: Qwen2.5-0.5B-Instruct

• —GRPO (Group Relative Policy Optimization)
• —150 training episodes
• —Multi-objective reward system

Efficiency:

• —QLoRA 4-bit (Unsloth)
• —Free T4 GPU
• —Training time: 81 minutes

📊 Results

👉 Improvement: +0.101 (+27%)

🔥 Key Discovery: Emergent Behavior

The model learned to protect the hospital:

• —Before: 31%
• —After: 89%
• —Improvement: +58%

We NEVER programmed this rule.

The model learned it through environment feedback and consequences.

🔗 Try It

• —🌐 Environment: https://huggingface.co/spaces/Ashith18/disasternet

• —🤖 Model: https://huggingface.co/Ashith18/sentinel-disasternet-grpo

• —💻 GitHub: https://github.com/Ashith-13/disasternet.git

🏁 Final Takeaway

GRPO training improved performance by +27% in 300 steps.

The model independently discovered hospital protection — a critical real-world strategy — without explicit programming.

[image].png)