Qwen2.5_Risk_1.5B — GGUF

GGUF quantized version of Firemedic15/Qwen2.5_Risk_1.5B, a Qwen2.5-1.5B-Instruct model fine-tuned on the Firemedic15/Travel_Risk_Data dataset for geopolitical travel risk analysis and OSINT-style security question answering.

This GGUF variant is intended for local inference via llama.cpp, Ollama, LM Studio, and compatible runtimes — no GPU required for most quantization levels.

---

Model Summary

Field	Value
Base model	Qwen2.5-1.5B-Instruct
Fine-tuned version	Firemedic15/Qwen2.5_Risk_1.5B
Format	GGUF (quantized)
Task	Travel risk Q&A, security threat assessment
Language	English
License	MIT

---

Intended Use

This model is built for security practitioners, travel risk analysts, and OSINT researchers who need fast, structured responses to questions like:

• What is the current conflict situation in [country]?
• What travel advisory level applies to [region]?
• What are the primary security risks for a traveler with [passport country] visiting [destination]?
• Summarize recent armed conflict events in [area] over the last 30 days.

It is deployed as the inference backend for the Firemedic15/OSINTTool Gradio Space, which combines this model with live data from ACLED, State Department advisories, RSS feeds, and airspace sources via a smolagents agentic loop.

This model is not suitable for:

• Medical, legal, or financial decisions
• Operational security planning without human verification
• Any context requiring real-time situational awareness (the model has a training cutoff and no live data access on its own)

---

Available Files

File	Quantization	Size (approx.)	Recommended For
Qwen2.5_Risk_1.5B.gguf	F16 (full precision)	~3.0 GB	GPU inference, maximum accuracy
Qwen2.5_Risk_1.5B-Q4_K_M.gguf	Q4_K_M	~0.9 GB	CPU inference, laptops, low RAM

Which should you use?

• If you have a GPU or want the highest fidelity output — use the F16 file.
• If you're running on a CPU, a laptop, or have less than 4 GB of RAM to spare — use Q4_K_M. For a 1.5B model the quality difference is minimal in practice.

---

Quickstart

Ollama

Option A — Pull directly from HuggingFace (no download needed):

ollama run hf.co/Firemedic15/Qwen2.5_Risk_1.5B

Option B — Install a local GGUF file:

This is useful if you've already downloaded the file or want to control which quantization you run.

Step 1 — Download the GGUF file

# F16 (full precision)
wget https://huggingface.co/Firemedic15/Qwen2.5_Risk_1.5B/resolve/main/Qwen2.5_Risk_1.5B.gguf

# Q4_K_M (quantized, recommended for CPU)
wget https://huggingface.co/Firemedic15/Qwen2.5_Risk_1.5B/resolve/main/Qwen2.5_Risk_1.5B-Q4_K_M.gguf

Step 2 — Create a Modelfile

Create a file named Modelfile (no extension) in the same directory as your GGUF:

FROM ./Qwen2.5_Risk_1.5B-Q4_K_M.gguf

SYSTEM """You are a professional OSINT intelligence analyst specializing in geopolitical conflict and security threat assessment. Provide structured, factual risk assessments based on the information available to you."""

PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 4096

Swap Qwen2.5_Risk_1.5B-Q4_K_M.gguf for the F16 filename if you downloaded that one instead.

Step 3 — Register the model with Ollama

ollama create qwen-risk -f Modelfile

Step 4 — Run it

ollama run qwen-risk

Once running, query it directly:

ollama run qwen-risk "What are the primary security risks for travelers in Sudan?"

Or use the OpenAI-compatible API that Ollama exposes:

import requests

response = requests.post(
    "http://localhost:11434/api/chat",
    json={
        "model": "qwen-risk",
        "messages": [
            {"role": "user", "content": "Assess the travel risk for Ukraine right now."}
        ],
        "stream": False,
    }
)
print(response.json()["message"]["content"])

llama.cpp (CLI)

# macOS
brew install llama.cpp
llama-cli -hf Firemedic15/Qwen2.5_Risk_1.5B

# Windows
winget install llama.cpp
llama-cli -hf Firemedic15/Qwen2.5_Risk_1.5B

llama.cpp (OpenAI-compatible server)

llama-server -hf Firemedic15/Qwen2.5_Risk_1.5B
# Exposes http://localhost:8080/v1 — compatible with any OpenAI client

llama-cpp-python

from llama_cpp import Llama

# F16 — full precision, best for GPU
llm = Llama.from_pretrained(
    repo_id="Firemedic15/Qwen2.5_Risk_1.5B",
    filename="Qwen2.5_Risk_1.5B.gguf",
)

# Q4_K_M — quantized, best for CPU / low RAM
llm = Llama.from_pretrained(
    repo_id="Firemedic15/Qwen2.5_Risk_1.5B",
    filename="Qwen2.5_Risk_1.5B-Q4_K_M.gguf",
)

response = llm.create_chat_completion(
    messages=[
        {
            "role": "system",
            "content": "You are a professional OSINT analyst specializing in geopolitical travel risk assessment."
        },
        {
            "role": "user",
            "content": "What are the primary security risks for travelers in Sudan right now?"
        }
    ],
    max_tokens=512,
    temperature=0.7,
)
print(response["choices"][0]["message"]["content"])

Docker

docker model run hf.co/Firemedic15/Qwen2.5_Risk_1.5B

LM Studio / Jan

Search for Firemedic15/Qwen2.5_Risk_1.5B in the model browser and download directly.

---

Prompt Format

This model uses the Qwen2.5 chat template. Structure your prompts as:

<|im_start|>system
You are a professional OSINT intelligence analyst specializing in geopolitical conflict and security threat assessment.
<|im_end|>
<|im_start|>user
[Your question here]
<|im_end|>
<|im_start|>assistant

For best results, include relevant context (country, timeframe, traveler profile) in the user turn rather than relying on the model's training data alone.

---

Training Details

• Base: Qwen2.5-1.5B-Instruct
• Method: Supervised fine-tuning (SFT) via HuggingFace TRL + PEFT
• Dataset: Firemedic15/Travel_Risk_Data
• Domain: Country-level risk assessments, conflict event summaries, advisory reasoning, embassy contact generation, airspace status interpretation
• Full precision version: Firemedic15/Qwen2.5_Risk_1.5B

---

Limitations

• No live data. The model has a training cutoff and does not fetch real-time conflict or advisory data on its own. For live analysis, use the OSINTTool Space which combines this model with live data pipelines.
• 1.5B scale. Tool-calling and structured JSON output are shallower than larger models. Outputs should be treated as analyst assistance, not authoritative assessments.
• English only. Fine-tuned on English-language data. Performance on non-English queries is not validated.
• Not for operational use without verification. AI-generated risk assessments from open sources are a starting point, not a finished product.

---

Related Resources

Resource	Link
Full precision model	Firemedic15/Qwen2.5_Risk_1.5B
Merged model	Firemedic15/qwen25-1.5B-travel-risk-analysis-merged
Training dataset	Firemedic15/Travel_Risk_Data
Live demo (Gradio Space)	Firemedic15/OSINTTool
GitHub	de0015

---

License

MIT. See LICENSE for full terms.

Built by Firemedic15 — security architect, former rescue paramedic, builder not buyer.