webshart/pseudo-camera-10k-structured

pseudo-camera-10k, native structured captions

The 9,997 images and structured captions from terminusresearch/pseudo-camera-10k-structured-json, pinned to revision 29b6406122ce01d82aa72148984e3c7ac7ae405e, repackaged into 19 indexed webshart tar shards. The original photographs come from bghira/pseudo-camera-10k. License metadata is inherited from the source dataset.

Image bytes are unchanged. Each source .txt caption is parsed into a native JSON object and embedded under files[filename].captions in its shard's JSON index. No caption objects are flattened or stored as escaped JSON strings. The tar members contain the original PNG images; dimensions and range offsets are indexed. Original sample paths such as train/00/<stem>.png are preserved.

Layout

data/shard-00000.tar
data/shard-00000.json
...
manifest.json

Shards target 1 GiB. manifest.json records the source revision, shard counts, and SHA-256 checksums. Every image was checked against the source LFS SHA-256, every caption object against its original JSON, and every indexed image dimension and byte range against the tar payload.

Native captions

The source schema is retained: high_level_description, style_description, and compositional_deconstruction, including nested elements and colour palettes. Bounding boxes remain in original image pixels. They are not normalized to 0–1000; Ideogram's caption verifier may report out-of-range coordinates, as it does for the original source captions. Caption descriptions and approximate boxes inherit the source dataset's limitations.

Use webshart, SimpleTuner, and CaptionFlow builds with native JSON caption support. Older webshart readers accepting only strings or lists of strings cannot read these caption objects.

import webshart

dataset = webshart.discover_dataset("webshart/pseudo-camera-10k-structured", subfolder="data")
loader = webshart.TarDataLoader(dataset, load_file_data=False)
entry = loader.load_sample(0, 0)
caption = entry.captions
print(caption["high_level_description"])

SimpleTuner

{
  "id": "pseudo-camera-10k-structured",
  "type": "webshart",
  "source": "webshart/pseudo-camera-10k-structured",
  "subfolder": "data",
  "caption_strategy": "webshart",
  "metadata_backend": "webshart",
  "resolution": 1024,
  "resolution_type": "pixel_area",
  "webshart": {"shard_cache_gb": 0}
}

Objects stay intact in metadata and caption caches. Each becomes one JSON text prompt when preparing text-encoder inputs; mixed lists of objects and strings remain separate caption alternatives. webshart.caption_key can select an individual field, such as high_level_description, for prose-only training.

CaptionFlow

Use the webshart export format with locally downloaded shard JSON indexes. CaptionFlow accepts native objects or mixed caption lists in Caption.outputs, and decodes JSON objects stored in Lance text columns when exporting. Its webshart export writes native values under the plural captions key.