Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

SARVIA: Models, data and benchmark checkpoints

Resources for SARVIA: Fast-Adapting the Strong SAR Foundation Models That Are Already in Your MLLMs, by Yuxuan Li, Zichen Tian and Qianru Sun.

Code and installation · arXiv:2611.22273 · PDF

SARVIA adapts the visual encoders of multimodal language models using self-supervised visual questions with computable answers and loss-adaptive task sampling. The released encoders are exported at step 600 and run without the language model.

All five bundles are subdirectories of this single dataset repository. Sizes are approximate decimal units; the complete release is about 1.46 TB.

Directory Size Contents
SARVIA-ViT-B 352 MB Base visual encoder, configuration and model card
SARVIA-ViT-L 1.23 GB Large visual encoder, configuration and model card
SARVIA-Benchmarks 11.66 GB Twelve final task checkpoints, configurations, metrics and logs
SARVIA-Pretrain 1.10 TB Three-source SAR corpus, metadata and task indexes
SARVIA-Optical 343.46 GB Paired optical assets and review images

Download

Run these commands from the cloned code repository. Download only the bundles needed for your task; the base encoder alone is sufficient for feature extraction.

python -m pip install -U huggingface_hub
hf download GreatBird/SARVIA --repo-type dataset \
  --include "SARVIA-ViT-B/*" --local-dir huggingface

Replace the include pattern with SARVIA-ViT-L/* for the large encoder, or select several bundles in one command:

# Both encoders and all downstream checkpoints (~13.24 GB).
hf download GreatBird/SARVIA --repo-type dataset \
  --include "SARVIA-ViT-B/*" "SARVIA-ViT-L/*" "SARVIA-Benchmarks/*" \
  --local-dir huggingface

# Pretraining corpus and paired optical assets (~1.44 TB).
hf download GreatBird/SARVIA --repo-type dataset \
  --include "SARVIA-Pretrain/*" "SARVIA-Optical/*" --local-dir huggingface

Keep these directory names so the training commands and local HTML example links resolve. sarvia.load_model accepts a local bundle directory; see the code README for the feature-extraction example and separate training environments.

Data and reproducibility

The SAR corpus contains 139,251 source samples: 119,251 SARLO pairs and 10,000 samples each from TerraMesh and KuroSiwo, including the documented training, validation and reserved splits. Human captions and semantic labels are excluded. See data documentation for source attribution, raw-data validity, geometry and split details.

The released subset reproduces the method and training settings. Its sample exposure differs from the original full-corpus runs, so retraining is not expected to produce identical weights or benchmark scores.

Licenses

The encoder and benchmark checkpoint bundles carry Apache-2.0 licenses. The SAR corpus follows its source-specific licenses. The optical companion is included with redistribution authorization confirmed by the project maintainers; its provider terms still apply. The code license does not relicense third-party imagery. party imagery.

Downloads last month
99

Paper for GreatBird/SARVIA