lfm-ucf-gguf

LFM2.5-VL-1.6B UCF Crime — GGUF

Base model: LiquidAI/LFM2.5-VL-1.6B fine-tuned on the UCF Crime dataset for surveillance crime detection.

Quantized GGUF files for fast inference with llama.cpp / Ollama / LM Studio — derived from rajofearth/lfm-ucf-unsloth.

🔓 Training notebook (free Colab): Open in Colab — see exactly how this model was trained and exported.

---

About this model

Fine-tuned using Unsloth on ~26k surveillance images from the UCF Crime dataset across 15 categories:

Abuse · Arrest · Arson · Assault · Burglary · Explosion · Fighting · Robbery · Shooting · Shoplifting · Stealing · Vandalism · Road Accident · Normal

The model analyzes surveillance images and outputs structured JSON:

{
  "isHarm": true,
  "descriptionIfHarm": "The image depicts a physical altercation."
}

When no harmful activity is detected:

{
  "isHarm": false
}

⚠️ Output format note: Always include a system prompt explicitly requesting JSON output — the model is trained toward it but won't default to that format without instruction.

---

Usage

llama.cpp

./llama-cli -m LFM2.5-VL-1.6B.Q4_K_M.gguf \
  --image your_surveillance_image.jpg \
  -p "Analyze this surveillance image and respond ONLY in JSON: {\"isHarm\": true/false, \"descriptionIfHarm\": \"reason if harmful, else omit\"}." \
  --temp 0.1

Ollama

# Create a Modelfile
cat > Modelfile << 'EOF'
FROM ./LFM2.5-VL-1.6B.Q4_K_M.gguf
SYSTEM "You are a surveillance analysis assistant. Analyze images for harmful or criminal activity. Always respond in strict JSON: {\"isHarm\": true/false, \"descriptionIfHarm\": \"brief description if harmful, else omit\"}."
EOF

ollama create lfm-ucf -f Modelfile
ollama run lfm-ucf

Note: Vision GGUF support for this architecture is still experimental in llama.cpp and Ollama. Results may vary — the LoRA version via Unsloth/Transformers is more reliable for production use.

---

Performance

+9.6 percentage point improvement on UCF Crime CCTV imagery. Evaluated using an LLM judge on a held-out test set.

---

Reproduce This Fine-Tune

The full pipeline — training, evaluation, and GGUF export — is available as a free public Colab notebook:

No paid GPU required. Runs on a free T4.

---

Related

---

Developed by: rajofearth · Created with Unsloth + Google Colab (free tier).