| --- |
| pretty_name: SARFACT |
| extra_gated_heading: Request access to SARFACT |
| extra_gated_description: Requests are reviewed manually. Your Hugging Face username, email, institution and intended use are shared with the maintainers to review and manage access. This information is not published in the dataset repository. Hidden Test is not distributed. |
| extra_gated_button_content: Submit access request |
| extra_gated_fields: |
| Institution or organization: text |
| Intended use: |
| type: select |
| options: |
| - Academic research |
| - Model development and benchmarking |
| - Teaching and coursework |
| - Industry research and applications |
| - Other |
| --- |
| |
| # SARFACT |
|
|
| A benchmark for grounded understanding of synthetic aperture radar imagery across nine tasks. |
|
|
| 1. Sign in to Hugging Face and submit the access request. |
| 2. Wait for approval from the maintainers. |
| 3. Return here and download the files with the same account. |
|
|
| | Partition | Images | Objects | Download | |
| |---|---:|---:|---| |
| | Train | 120,208 | 675,420 | [Files](https://huggingface.co/datasets/SARFACT-Benchmark/SARFACT/tree/main/train) | |
| | Validation | 20,042 | 105,372 | [Files](https://huggingface.co/datasets/SARFACT-Benchmark/SARFACT/tree/main/val) | |
| | Public Test | 30,059 | 164,905 | [Files](https://huggingface.co/datasets/SARFACT-Benchmark/SARFACT/tree/main/test) | |
| | Hidden Test | — | — | Internal reserve — not distributed | |
|
|
| Access is granted to individual Hugging Face accounts. Approval does not grant permission to redistribute the dataset or override source-specific usage conditions. |
|
|
| ## Download and extract |
|
|
| Each partition has an `annotations.tar` and independently extractable `images-*.tar` shards. Download **all** image shards for the selected partition. These are separate tar archives, not pieces to concatenate. Extract all archives into the same dataset directory. |
|
|
| | Partition | Image archives | Annotation archives | |
| |---|---:|---:| |
| | Train | 11 | 1 | |
| | Validation | 2 | 1 | |
| | Public Test | 3 | 1 | |
|
|
| Image archives are approximately 1 GiB each, except the final shard of each partition. Images retain their original bytes; packaging does not recompress or alter them. |
|
|
| ```bash |
| pip install -U huggingface_hub |
| hf auth login |
| hf download SARFACT-Benchmark/SARFACT --repo-type dataset --include "val/*" "SHA256SUMS" --local-dir downloads |
| (cd downloads && sha256sum --ignore-missing -c SHA256SUMS) |
| mkdir -p SARFACT |
| for archive in downloads/val/*.tar; do tar -xf "$archive" -C SARFACT; done |
| ``` |
|
|
| Replace `val` with `train` or `test` as needed. Use your own approved HF account. If interrupted, rerun the download command with the same destination. Confirm that all archives listed for your partition are present, and that their SHA256 checks pass before extraction. `--ignore-missing` skips partitions you did not download; it does not check that your selected partition is complete. |
|
|
| ## Data layout |
|
|
| ```text |
| SARFACT/ |
| images/<split>/* |
| object_pool/<split>.jsonl.gz |
| tasks/<task>/<split>.jsonl.gz |
| ``` |
|
|
| - Each gzip JSONL line is one record. Read with `gzip.open(path, "rt")` and `json.loads(line)`. |
| - Resolve each record's `image` relative to the `SARFACT/` directory. |
| - Object-pool records contain image dimensions, object IDs, foreground categories, boxes and certified absent categories. |
| - Task records contain a query, answer, image reference and task-specific fields. Where applicable, `evidence` refers to object IDs, categories, roles and HBBs. |
| - HBBs use native image pixels in `[xmin, ymin, xmax, ymax]` order. Decimal coordinates are intentional; auxiliary polygons do not change the HBB output contract. |
| - Task directories: `detection`, `presence`, `classification`, `count`, `point_conditioned_recognition`, `spatial_relation`, `comparison`, `referring_expression` (Visual Grounding), `caption`. |
| - Use the provided splits. Hidden Test images, queries, annotations and transformation mappings are excluded. |
|
|
| ## Sources and use |
|
|
| See [SOURCES_AND_LICENSES.md](SOURCES_AND_LICENSES.md) for upstream attribution and usage restrictions. Access approval does not override those restrictions. |
|
|
| ## Point-coordinate update (2026-11-01) |
|
|
| Point queries and `point` fields use integer original-image pixel coordinates. Points uniquely identify their target in native coordinates and after the benchmark's normalized-coordinate roundtrip. Two Train targets were replaced by other objects of the same category in the same images; a two-record Validation group was removed because one target had no valid integer point. Validation Point now contains 1,725 records; Public Test remains at 2,700. Non-Point tasks, images, and object-pool bounding boxes are unchanged. |
|
|
