LFM2.5-VL-450M-landslide-GGUF

LFM2.5-VL-450M — Landslide Risk (GGUF)

Fine-tuned LiquidAI/LFM2-VL-450M for landslide hazard / exposure risk assessment from seven-band Sentinel-2 composites. Distributed as GGUF for llama.cpp / llama-server inference.

Files

Use both files together — --mmproj is required for image input.

Inference (llama-server)

llama-server \
  -m landslide-Q8_0.gguf \
  --mmproj mmproj-landslide.gguf \
  --jinja --port 8080

Then call the OpenAI-compatible chat endpoint at http://127.0.0.1:8080/v1/chat/completions with seven Sentinel-2 composite images per sample (RGB, SWIR, NIR, NDMI, NDVI, BSI, TCT Wetness — in that exact order).

Input format

Per sample the model expects:

• Seven Sentinel-2 imagery composites (PNG)
• Scene metadata: longitude, latitude, landslide movement type, capture month, cloud cover
• The fine-tuning system prompt (see Sciamlab/landslide-sentinel2-vlm)

Output

A single JSON object with these fields (all floats 1.0–4.0 unless noted):

soil_moisture_anomaly, vegetation_stress, bare_soil_exposure, water_accumulation, steep_terrain_visible, erosion_evidence, swir_anomaly, tct_wetness_anomaly, image_quality_limited (bool), exposure_risk, exposure_risk_explanation, geo_features (list), hazard_risk, hazard_risk_explanation.

Post-inference: overall_risk = (hazard_risk × exposure_risk − 1) / 15.

Training

• Base: LiquidAI/LFM2-VL-450M
• Framework: leap-finetune
• Recipe: Full VLM SFT (no LoRA), 3 epochs, lr 2e-5 cosine, warmup 0.03
• Hardware: 1× H100 (Modal), ~31 min wall clock
• Eval loss @ epoch 1: 1.015

Pipeline source and prompts: github.com/Sciamlab/landslide-prevention (cookbook fork of Liquid4All/cookbook/examples/wildfire-prevention).

Quantization

Converted with llama.cpp pinned at commit db44417 (pre-#22597, which broke Lfm2Vl --mmproj extraction with Can not map tensor 'lm_head.weight').

Limitations

• Trained on a curated landslide subset; performance on out-of-distribution scenes (deserts, ice, dense urban) not characterized.
• Q8_0 backbone — minor quality loss vs F16. Expect modest field-accuracy drop.
• Vision tower mostly unchanged from the base model; classification-level errors on edge cases inherit from the base.

Citation

If you find the model & data useful, please cite:

@software{Sciamlab/LFM2.5-VL-450M-landslide-GGUF,
  author = {Sameer Ahmed, Alessio Dragoni, Tommaso Biondo},
  title = {LFM2.5-VL-450M — Landslide Risk (GGUF): Vision Model for Landslide Early Warning on edge},
  month = {May},
  year = {2026},
  url = {https://huggingface.co/Sciamlab/LFM2.5-VL-450M-landslide-GGUF}
}