FIRM

FIRM (Fine-Grained Intra-Token Representation of Masks) represents each compressed visual token with a 2x2 binary sub-cell mask code and refines the resulting structural field with a lightweight continuous renderer for fine-grained remote-sensing reasoning segmentation.

Configurations

The release provides one Qwen3-VL-4B default configuration for each dataset:

• configs/lasers.yaml
• configs/drseg.yaml
• configs/earthreason.yaml
• configs/risbench.yaml
• configs/rrsis.yaml

All dataset configurations inherit the shared settings in configs/default.yaml.

Local Paths

Run commands from the repository root. Place the model and datasets under the configured relative paths:

models/Qwen3-VL-4B-Instruct
data/LaSeRS
data/DRSeg
data/EarthReason
data/RISBench_dataset
data/RRSIS-D/rrsis/RRSIS-D

Update the relative paths or GPU IDs in the YAML files when using a different local layout.

Train

python scripts/train.py --config configs/lasers.yaml

Evaluate

python scripts/eval.py --config configs/lasers.yaml --checkpoint outputs/run_name --split all