lfm2-flood

Fine-tune of LiquidAI/LFM2.5-VL-450M for flood detection from Sentinel-2 satellite tile pairs (RGB + SWIR baseline + RGB + SWIR current). Output is a structured JSON flood-risk profile.

The model is intended to run on a low-resource node (a satellite or a community ground station) and downlink only the JSON payload — not raw imagery — for offline humanitarian response in flood-affected regions.

Status: experimental. Trained on 88 paired samples (110 total, 22 held out for eval). The dataset is too small and noisy for an operational alert system; this is the infrastructure, not a deployable model. See the Limitations section.

Inputs

Four images per inference, in this order:

1. RGB-baseline (B4-B3-B2 true color) — the location at a pre-event timestamp
2. SWIR-baseline (B12-B8-B4 false color) — same timestamp
3. RGB-current — the location at the timestamp to assess
4. SWIR-current — same timestamp

Each is a Sentinel-2 PNG, ~5 km × 5 km, served as base64 data URL or HF Hub URL.

Output

{
  "flood_present": true,
  "flood_severity": "moderate",
  "water_coverage_pct_estimate": "30-60%",
  "populated_area_affected": true,
  "infrastructure_at_risk": true,
  "river_overflow_visible": true,
  "image_quality_limited": false
}

7 fields. flood_present and the four booleans are change-relative — i.e. "is there flooding vs the baseline tile", not "is there water in this tile". image_quality_limited is the abstention signal — set true on cloudy or partial-coverage tiles.

Files

file	size	purpose
model.safetensors	~860 MB	merged HF checkpoint, full fine-tune of LFM2.5-VL-450M
config.json, tokenizer*.json, chat_template.jinja, etc.	small	standard HF transformers metadata
lfm2-flood-Q4_0.gguf	~245 MB	quantized backbone for llama.cpp
mmproj-lfm2-flood-F16.gguf	~189 MB	vision tower + projector for llama.cpp

Quick start (llama.cpp)

# Download both files (or use `hf download`)
hf download jpmarindiaz/lfm2-flood lfm2-flood-Q4_0.gguf --local-dir .
hf download jpmarindiaz/lfm2-flood mmproj-lfm2-flood-F16.gguf --local-dir .

# Serve OpenAI-compatible endpoint
llama-server -m lfm2-flood-Q4_0.gguf --mmproj mmproj-lfm2-flood-F16.gguf -c 8192 --port 8765

Then POST 4 images to /v1/chat/completions with a JSON-schema response_format. Full client code in the humaid repo (finetune-flood/src/evaluate.ts, finetune-flood/app/server.ts).

Training

• Base model: LiquidAI/LFM2.5-VL-450M
• Backend: leap-finetune on Modal H100
• Training type: vlm_sft, full fine-tune (no LoRA — the multimodal projector needs to relearn satellite multispectral imagery)
• Hyperparameters: 3 epochs, per-device batch 2, gradient accumulation 8 (effective 16), LR 5e-5 cosine, vision-encoder LR multiplier 0.5, bf16
• Wall time: ~70 seconds on H100
• Dataset: 88 train + 22 eval pair samples from 9 documented La Mojana / Putumayo flood events. See jpmarindiaz/flood-detection-pair-colombia.

Evaluation

Compared against the labeler's self-consistency oracle and the un-fine-tuned base model on the same 110 pair samples:

field	opus oracle (n=30)	base LFM2.5-VL (n=110)	this model (n=110)
valid_json	1.00	1.00	1.00
fields_present	1.00	1.00	1.00
flood_present	0.67	0.66	0.66
flood_severity	0.43	0.29	0.29
water_coverage_pct_estimate	0.70	0.37	0.35
populated_area_affected	0.73	0.51	0.51
infrastructure_at_risk	0.73	0.54	0.54
river_overflow_visible	0.67	0.60	0.60
image_quality_limited	0.83	0.10	0.90
overall	0.68	0.44	0.55
avg latency (s)	3.87	0.53	0.53

overall is the macro-average across the 7 fields. valid_json and fields_present are 1.0 because we use grammar-constrained JSON output via response_format: {type: "json_schema"}.

Limitations

1. Sentinel-2 wet-season cloud cover. ~50% of acquisitions over La Mojana in Apr–Jun and Aug–Nov are >50% cloud. Operational pipelines (CopernicusLAC) use Sentinel-1 SAR (cloud-independent) for the same task. This model was trained only on Sentinel-2.
2. Inter-labeler noise floor caps the ceiling. Opus self-consistency on the hardest schema fields is 0.43–0.70. The student model can't exceed inter-labeler agreement.
3. The fine-tune learned the dataset prior, not the task. The +11-point overall gain (0.44 → 0.55) is concentrated almost entirely in image_quality_limited (0.10 → 0.90). Other fields essentially didn't move with 88 train samples × 3 epochs.
4. Geographic scope: Colombia only. Locations are 8 La Mojana municipalities + 6 Putumayo. May not generalize to other flood regimes (e.g. flash floods, glacial-melt rivers, hurricane storm surge).

How it was built

The full pipeline + scripts + playbook are in the humaid repo on GitHub. Key files:

• finetune-flood/PLAYBOOK.md — end-to-end command sequence (fetch → label → build → upload → train → pull → package → serve → eval)
• finetune-flood/REPORT.md — wrap-up findings (what worked, what didn't, the case for switching to Sentinel-1 SAR before resuming)
• finetune-flood/docs/ — overview, pipeline, data collection, labeling, evaluation, findings
• finetune-flood/scripts/convert_mmproj_lfm2vl.py — patch around the upstream convert_hf_to_gguf.py issue with lm_head.weight in full-FT merged checkpoints
• finetune-flood/app/ — small Hono app for testing the model interactively

Citation

Trained 2026-05-08 as part of the humaid project — offline-first humanitarian response toolkit for flood crises in Colombia. Built on:

• LFM2.5-VL-450M by Liquid AI
• leap-finetune by Liquid AI
• SimSat by DPhi Space
• The Liquid AI × DPhi Space AI in Space hackathon wildfire-prevention example by Pau Labarta Bajo