AI Or Die — Wilderness Survival AI

Offline · Agentic · Context-Aware

📊 Presentation Slides · 🤗 Live Demo

Lost in the wilderness with no signal? One wrong mushroom can kill you in 6 hours. AI Or Die identifies potentially lethal wild species from a photo and issues a survival verdict — no internet required.

Wi-Fi covers only ~20% of the Earth's land surface. Emergencies happen in the other 80%.

Built at Mistral AI Hackathon 2025.

---

How It Works

Photo + Location Context  (image · GPS · season · environment)
              ↓
   Fine-tuned Ministral-3B  ←→  Local species_db (offline JSON)
         [TOOL_CALLS]
              ↓
   🟢 SAFE  /  🔴 DO NOT EAT  /  🚨 EMERGENCY

The model is trained not to memorize species facts, but to call the right tool, interpret the DB result, and issue the correct warning. Facts live in a structured local DB — zero hallucination.

---

Key Design Choices

Approach	Problem
Fine-tune for knowledge	Hallucination — model invents facts
Cloud API (GPT-4, Gemini)	Needs internet — fails in the wilderness
Fine-tune for agentic behavior + local DB	✅ Reliable facts. ✅ Works offline.

---

Model

• Base: Ministral-3B
• Method: QLoRA (4-bit NF4), r=32, alpha=32
• Trained on: 1,394 agentic tool-calling conversations
• HF repo: Yongtae723/survive-or-die-lora

Tools the model learns to call

species_db_lookup(species_guess, category, confidence)
emergency_protocol(species_guess, category, time_since_ingestion)
nearby_species_search(latitude, month, altitude_m, environment)

---

Training Data Pipeline

iNaturalist (7,900+ images)     Google Gemini 2.5 Flash
        ↓                               ↓
  5 mushroom species              species_db.json
  (one per safety category)       (toxins · habitat · range)
        ↓
  Rule Engine (verdict logic) + LLM (natural language)
        ↓
  train.jsonl — 1,394 samples (1,172 / 108 / 114)
        ↓
  Ministral-3B Fine-tuning

Species used for training:

Species	Category
Amanita phalloides	LETHAL (Death Cap)
Amanita muscaria	DO_NOT_EAT (Fly Agaric)
Flammulina velutipes	CONFUSING (Enoki — wild has lethal lookalikes)
Lentinula edodes	CONDITIONAL_SAFE (Shiitake)
Hericium erinaceus	SAFE (Lion's Mane)

---

Evaluation

3-way comparison (Gemini 2.5 Flash Lite vs Base Ministral-3B vs Fine-tuned) on 75 test samples via W&B Weave:

Metric	Gemini Flash Lite	Base Ministral-3B	Fine-tuned
Safety recall (lethal cases)	moderate	low	↑ highest
Species exact match	moderate	low	↑ highest
Tool call accuracy	n/a	n/a	~high

---

Project Structure

├── src/
│   ├── app/                        # Gradio demo app (HF Spaces)
│   ├── data-prep/                  # Training data pipeline
│   ├── train/                      # QLoRA fine-tuning
│   └── eval/                       # Weave 3-way evaluation
├── notebooks/                      # Colab inference + eval notebook
└── requirements.txt

---

Safety Principle

"If in doubt, don't eat."

• Lethal species → absolute refusal, no exceptions
• Species with dangerous lookalikes → DO NOT EAT even if edible
• Unknown species → DO NOT EAT
• Emergency ingestion → skip identification, trigger emergency protocol immediately