--- language: - en base_model: Qwen/Qwen3.5-2B pipeline_tag: image-segmentation tags: - remote-sensing - referring-segmentation - reasoning-segmentation - lora - georevis --- # GeoReVis **GeoReVis: Learning Selective Revision and Evidence Consolidation for Remote Sensing Segmentation** This repository contains the final GeoReVis weights trained on the full eligible LaSeRS training set. GeoReVis combines execution-feedback revision, Candidate Selection, and Cross-Mask Evidence Consolidation over three rounds. The model uses Qwen3.5-2B with role-specific LoRA adapters and a frozen SAM3 segmentation decoder. Its final LaSeRS paper results are **72.00% gIoU** and **77.01% cIoU**. ## Files | Directory | Role | |---|---| | `planner/` | Target Decomposition adapter | | `initializer/` | Merged Spatial Grounding model and processor | | `reviser/` | Execution-Verified Feedback Revision adapter | | `selector/` | Candidate Selection adapter | | `critic/` | Feedback Assessment adapter | | `feedback_reviser/` | Multi-round Feedback Revision adapter | | `transition_selector/` | Multi-round Candidate Selection adapter | | `consolidation/adapter/` | Final Cross-Mask Evidence Consolidation adapter | | `consolidation/readout.pt` | Joint regional readout, including its frozen initialization branch | | `inference.json` | Portable configuration for the GeoReVis code release | | `SHA256SUMS` | File checksums | All listed roles are used in the final system. The consolidation adapter alone is not a standalone segmentation model. Training checkpoints, optimizer states, logs, and evaluation outputs are not included. ## Usage Use these weights with the **GeoReVis_OpenSource** code release. Download this repository into `GeoReVis_FinalWeights/` beside the code directory: ```bash hf download georevis/GeoReVis --local-dir GeoReVis_FinalWeights ``` ```text workspace/ GeoReVis_OpenSource/ GeoReVis_FinalWeights/ ``` Follow the code README to install the environment and prepare the external [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) backbone, [SAM3](https://huggingface.co/facebook/sam3), and TRAIN visual reference library. These external assets are not included in this repository. The reference library is constructed by the Execution-Driven Data Pipeline. Run the code's inference launcher with `--config ../GeoReVis_FinalWeights/inference.json`. The configuration resolves `${project}` to the code checkout and specifies all role weights explicitly. ```bash cd GeoReVis_OpenSource python scripts/run_inference.py \ --config ../GeoReVis_FinalWeights/inference.json --gpus 0,1,2,3 \ --inputs datasets/manifests/lasers/test/inputs.jsonl \ --output outputs/pretrained/lasers_test ``` The Execution-Driven Data Pipeline uses **MiniMax-M3.1** to improve training-data quality. It is not needed at inference once the trained weights and reference library are available. ## Integrity and Base Model Terms To verify downloaded files, run `sha256sum -c SHA256SUMS` in the weights directory. The Qwen3.5-2B base-model license is retained in `LICENSE-Qwen3.5.txt`. SAM3 and datasets retain their own distribution terms.