Geo

# Remote Sensing Multitask SFT Multimodal supervised fine-tuning data for remote-sensing tasks. > **Release status:** keep this repository private until redistribution rights > for every upstream image dataset have been reviewed and documented. ## Dataset summary - Records: 19,009 - Unique images: 13,923 - Image references: 19,159 - Logical image size: 173.76 GiB - Format: JSON Lines with ModelScope-SWIFT style `messages` and `images` ## Tasks | Task code | Records | | --- | ---: | | CCR | 1,247 | | OCC | 2,355 | | OCL | 2,354 | | OLUC | 5,000 | | OMS | 291 | | OSR | 4,080 | | RC | 900 | | RCCD | 150 | | RLUC | 800 | | RP | 1,832 | ## Image provenance groups | Source group | Unique images | | --- | ---: | | DOTA-v1.0 | 623 | | FAIR1M | 1,551 | | GeoLLaVA-Data | 4,153 | | OLUC | 5,000 | | RRSIS | 200 | | SpaceNet3-derived | 1,832 | | xView | 564 | The source-group counts describe file provenance only. They do not grant or replace the licenses of the upstream datasets. ## Repository layout ```text . ├── README.md ├── train.jsonl ├── validation_report.json └── images/ ├── 000/ ├── 001/ └── ... ``` Every value in `images` is a path relative to the repository root, for example: ```json {"images":["images/000/00000000.jpg"]} ``` Machine-local source paths were removed from the published JSONL. ## Download ```bash ms-hub download <namespace>/remote-sensing-multitask-sft --repo-type dataset --local-dir ./remote_sensing_multitask_sft ``` The legacy ModelScope CLI can also be used: ```bash modelscope download --dataset <namespace>/remote-sensing-multitask-sft --local_dir ./remote_sensing_multitask_sft ``` Run consumers from the downloaded repository root so that relative image paths resolve correctly. ## License and attribution This dataset combines annotations and imagery derived from multiple upstream remote-sensing datasets, including GeoLLaVA-Data, FAIR1M, SpaceNet3, xView, DOTA-v1.0, RRSIS, and fMoW/OLUC-related data. Before making this repository public, document the exact version, license, redistribution terms, and citation for every source. Where redistribution is not permitted, remove those images and provide a reconstruction/download script instead.