LoRA adapter for LiquidAI/LFM2.5-VL-1.6B, fine-tuned for satellite change detection across 7 event schemas (construction, flood, ice, thermal, urban, vegetation, wetland).

Eval

On v2_val (700 rows, 33 positives + 67 nulls per schema):

metric	value
F1	0.95
accuracy	0.969
precision	0.98
recall	0.92

Per-schema F1: IceExtent 1.00, ThermalHotspot 0.97, VegetationStress 0.97, UrbanChange 0.95, WetlandChange 0.95, FloodExtent 0.91, ConstructionProgress 0.90.

Null-baseline accuracy on this set is 0.670.

Recipe

Base: LiquidAI/LFM2.5-VL-1.6B (bf16)
LoRA: r=16, alpha=16, dropout=0, targets q_proj k_proj v_proj out_proj in_proj w1 w2 w3
Vision frozen (finetune_vision_layers=False); language tower only
4 epochs × 29,281 rows (full MOSAIC TRAIN split, natural class distribution), effective batch 8, lr 2e-4 linear, weight_decay 0.001
H100 80GB, ~5h wall

Inference

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoProcessor
import torch

base = AutoModelForCausalLM.from_pretrained(
    "LiquidAI/LFM2.5-VL-1.6B",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).eval()
model = PeftModel.from_pretrained(base, "danvinci/lfm25vl_mosaic_change_detection")
processor = AutoProcessor.from_pretrained("LiquidAI/LFM2.5-VL-1.6B", trust_remote_code=True)

Generation defaults: do_sample=True, temperature=0.1, min_p=0.15, repetition_penalty=1.05 (Liquid's published values). Greedy decoding underperforms.

Prompt contract

Per-schema predicate prompt of the form "did <schema> happen?" with two image inputs (before, after). The model returns either a JSON object matching the schema's field spec or the literal string null. See training/prompts.py for predicate_prompt() and parse_output().

Source

https://github.com/danvinci/liquid_space