metadata
license: other
task_categories:
- image-text-to-text
- image-segmentation
language:
- en
pretty_name: PixVL Cross-Dataset Self-Supervised Data
PixVL Cross-Dataset Self-Supervised Data
Synthetic self-supervised winner/preference records generated for PixVL stage1 training.
Download
The release is stored as split tar parts:
cat pixvl_cross_dataset_selfsup_stage1_422032_20260515.tar.part-* > pixvl_cross_dataset_selfsup_stage1_422032_20260515.tar
sha256sum -c pixvl_cross_dataset_selfsup_stage1_422032_20260515.tar.sha256
tar -xf pixvl_cross_dataset_selfsup_stage1_422032_20260515.tar
Expected tar SHA256:
825c0b40267d4d949132f62800ee5aac0471ad900566f23405b252f34ef37300
Contents After Extraction
merged/merged_unsup_unique.jsonl: winner-only training records, 422,032 lines.merged/merged_preference_cache.jsonl: preference cache records, 124,689 lines.merged/merge_stats.json: merge metadata.sav_chooseone/images/: generated SAV source/target images.coco_chooseone/images/: generated COCO source/target images.cityscapes_chooseone/images/: generated Cityscapes source/target images.openimages_chooseone/images/: generated OpenImages source/target images.
The image field in merged/merged_unsup_unique.jsonl is relative to the extracted dataset root.
Source Counts
- SAV: 410,440 records
- COCO: 5,780 records
- Cityscapes: 1,724 records
- OpenImages: 4,092 records
- Total winner records: 422,032
Export Note
Two SAV pairs were excluded from this export after transient filesystem I/O errors during packaging. The final winner count is 422,032 records.