Offline survival assistant fine-tuned from Qwen/Qwen2.5-0.5B-Instruct on a curated corpus of survival, first-aid, foraging, navigation, and engineering reference material. Designed to run fully offline on phones (iPhone SE 2 / Android with 3GB+ RAM) via PocketPal AI, llama.cpp, or any GGUF-compatible runtime.
Model Files
File
Size
Use
v3-survival-qwen-0.5b-q4_k_m.gguf
380 MB
Recommended — latest fine-tune, best quality/size tradeoff
survival-qwen-0.5b-q4_k_m.gguf
380 MB
v1 fine-tune (older)
qwen2.5-0.5b-instruct-q4_k_m.gguf
380 MB
Base Qwen 2.5 0.5B (no fine-tune) for comparison
Quick Start — PocketPal AI on iOS
Open PocketPal AI → Models → + Hugging Face
Search: wattsonandme/survival-qwen-0.5b
Download v3-survival-qwen-0.5b-q4_k_m.gguf
Apply settings below
Recommended Inference Settings
These settings are critical — defaults will produce repetitive or off-topic output on this small model.
Setting
Value
Chat template
chatml (auto-loaded from GGUF)
Temperature
0.7
Top-P
0.9
Repeat penalty
1.15
Repeat last N
64
n_predict
512
Context length
2048
Recommended System Prompt
You are SurvivalAI, a comprehensive offline survival and rebuilding assistant. You provide practical, step-by-step guidance for: emergency medicine and first aid (injuries, illness, medications, dosages); water purification, food foraging, and agriculture; shelter construction and fire-starting; navigation by stars, sun, and landmarks; electrical systems, generators, radio communication, and basic engineering; nuclear/chemical/biological emergency response; animal and plant identification (edible vs toxic); tool making, repair, and improvised solutions; government structure, law, and community organization; seasonal almanac: planting times, weather patterns, tides. Be specific with quantities, temperatures, times, and measurements. If unsure, say so.
Coverage
The fine-tuning corpus covers:
First aid & medicine — Special Forces Medical Handbook, Where There Is No Doctor, FM 4-25.11, NAVMED references
Water — FM 21-10 Field Hygiene & Sanitation, EPA Water Purification
Shelter & fire — FM 31-70 Cold Weather, FEMA fallout shelter guides
Navigation — FM 3-25.26 Map Reading & Land Navigation
Military survival — FM 21-76, FM 21-76-1 SERE, USMC Survival
Disaster response — FEMA Citizen Guide, Nuclear War Survival Skills
Honest Caveats
0.5B is a small model. It will sometimes hallucinate dosages, plant IDs, or technical details. Verify life-critical answers against a primary source (PDF library, Kiwix offline Wikipedia, etc.).
For higher reliability, pair this model with a RAG pipeline retrieving from a vetted survival corpus on-device.
Tested on: iPhone SE 2 (A13, 3GB RAM, iOS 18) via PocketPal AI — runs at ~10-15 tok/s.
Why HuggingFace's Web Inference Looks Broken
The HF inference widget defaults to repetition_penalty=1.0 and may not apply the system prompt. On a 0.5B fine-tune, that produces looping or off-topic output. Use PocketPal or llama.cpp with the settings above for actual quality assessment.
Quantization: Q4_K_M via llama.cpp convert_hf_to_gguf.py + llama-quantize
License
Apache 2.0 (inherited from Qwen 2.5 base model). Training corpus consists of public-domain US military field manuals, Hesperian Foundation creative-commons medical references, and FEMA public-domain guides.
Project
Part of the Pantheon AI / SurvivalAI offline survival device project — an iPhone SE 2 dedicated entirely to offline survival reference (LLM + Kiwix + 70+ survival PDFs in Apple Books).
