LFM2.5-VL-450M wildfire risk (GGUF)

Fine-tuned from LiquidAI/LFM2.5-VL-450M on Sentinel-2 satellite imagery to assess wildfire risk. Part of the Liquid Cookbook wildfire-prevention example.

Given an RGB and SWIR Sentinel-2 image pair, the model outputs a structured JSON risk assessment:

{
  "risk_level": "low | medium | high",
  "dry_vegetation_present": true,
  "urban_interface": false,
  "steep_terrain": true,
  "water_body_present": false,
  "image_quality_limited": false
}

Eval results

Evaluated on 172 test samples from Paulescu/wildfire-prevention, ground truth from claude-opus-4-6.

field	claude-opus-4-6	LFM2.5-VL-450M Q8_0 (base)	LFM2.5-VL-450M Q8_0 (fine-tuned)
valid_json	1.00	1.00	1.00
fields_present	1.00	1.00	1.00
risk_level	0.99	0.08	0.76
dry_vegetation_present	0.99	0.48	0.83
urban_interface	0.98	0.25	0.93
steep_terrain	0.99	0.45	0.81
water_body_present	0.99	0.74	0.87
image_quality_limited	1.00	0.28	0.86
overall	0.99	0.38	0.84
avg latency (s)	2.91	0.72	0.59

Files

Running inference with a VLM in llama.cpp requires two GGUF files:

file	description
`lfm2.5-vl-wildfire-Q8_0.gguf`	Language model backbone (Q8_0)
`mmproj-lfm2.5-vl-wildfire-Q8_0.gguf`	Vision tower + multimodal projector (F16)

Usage

llama-server

llama-server \
    -m lfm2.5-vl-wildfire-Q8_0.gguf \
    --mmproj mmproj-lfm2.5-vl-wildfire-Q8_0.gguf \
    --jinja --port 8080

Reproduce eval results

Clone the Liquid Cookbook, then:

cd examples/wildfire-prevention
uv sync
uv run scripts/evaluate.py \
    --hf-dataset Paulescu/wildfire-prevention \
    --backend local \
    --model lfm2.5-vl-wildfire-Q8_0.gguf \
    --mmproj mmproj-lfm2.5-vl-wildfire-Q8_0.gguf \
    --split test