---
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
datasets:
- i-am-mushfiq/FirstAidQA
language:
- en
pipeline_tag: text-generation
tags:
- first-aid
- emergency
- offline-ai
- gguf
- llama-cpp
- qwen2
- medical-assistance
---

# Atlas-FirstAid (Qwen2.5-1.5B-Instruct Fine-Tune)

**Atlas** is a lightweight, edge-optimized language model fine-tuned specifically for first-aid and emergency guidance in offline, disconnected, or low-resource environments. 

It powers **[Nova F-R (Nova First Response)](https://github.com/Neuille-hush/NovaFR)**—a privacy-first mobile app designed to provide clear, step-by-step medical instructions during disasters, infrastructure outages, or wilderness emergencies when internet access is completely unavailable.

---

## 📌 Model Details

- **Developer:** Ismet Beljulji
- **Base Model:** `Qwen/Qwen2.5-1.5B-Instruct`
- **Fine-Tuning Framework:** Unsloth
- **Dataset:** [FirstAidQA](https://huggingface.co/datasets/i-am-mushfiq/FirstAidQA) (5,500 question-answer pairs derived from the certified *Vital First Aid Book*, presented at the NeurIPS 2025 Workshop on Muslims in ML)
- **Primary Quantization:** GGUF (`Q4_K_M`, ~986 MB)
- **Target Execution Engine:** `llama.cpp` on mobile/edge hardware

---

## ⚙️ Recommended System Prompt

For optimal response structure and safety, pass the following system prompt to the engine:

```text
You are Atlas, an AI emergency first-aid assistant trained to provide immediate, calm, and actionable guidance during health crises. 

Your instructions must be:
1. Clear, numbered, step-by-step, and prioritized by life safety (e.g., checking breathing/airway first).
2. Direct and concise—avoid unnecessary conversational filler in an emergency.
3. Explicit about when to seek immediate emergency services (call 911 / 112) when available.
4. Grounded strictly in standard first-aid protocols. Do not fabricate medical treatments.