webshart/conceptual-captions-12m-webdataset-metadata

Conceptual Captions 12M — Webshart metadata indices

Per-shard webshart metadata indices for laion/conceptual-captions-12m-webdataset: 1,100 JSON files under data/, one per source tar shard, mirroring the source's shard layout.

Each index records every tar member's byte offset and length (enabling ranged reads without downloading whole shards), image geometry (width/height for aspect bucketing), and — as of August 2026 — embedded captions for all 10,994,853 samples, coalesced from the dataset's .txt sidecar members with webshart optimize-captions. Consumers no longer need per-sample range reads to fetch captions; they arrive with the metadata.

Licensing of the underlying images/captions follows the source dataset; this repository only contains derived index metadata.

Usage with the webshart loader

import webshart

dataset = webshart.discover_dataset(
    source="laion/conceptual-captions-12m-webdataset",
    metadata="webshart/conceptual-captions-12m-webdataset-metadata",
    hf_token=None,  # or a token / HF_TOKEN env for gated access
)
dataset.enable_metadata_cache(location="cache/metadata")
dataset.enable_shard_cache(location="cache/shards", cache_limit_gb=25, parallel_downloads=4)

loader = webshart.TarDataLoader(dataset, load_file_data=True)

# Captions come straight from the index:
entry = loader.load_sample(0, 0)
print(entry.caption)          # first caption string
metadata = loader.get_metadata(0)
print(metadata["00000000.jpg"]["captions"])

# Aspect bucketing for training pipelines:
buckets = loader.list_shard_sample_aspect_buckets(
    [0], key="aspect", target_pixel_area=1024 * 1024, target_resolution_multiple=64
)

Usage with SimpleTuner

A multidatabackend.json entry streaming this dataset with cached VAE latents and webshart-sourced captions:

[
  {
    "id": "cc12m-webshart-1024",
    "type": "webshart",
    "dataset_type": "image",
    "source": "laion/conceptual-captions-12m-webdataset",
    "metadata": "webshart/conceptual-captions-12m-webdataset-metadata",
    "caption_strategy": "webshart",
    "metadata_backend": "webshart",
    "crop": true,
    "crop_style": "random",
    "crop_aspect": "square",
    "minimum_image_size": 512,
    "maximum_image_size": 1024,
    "target_downsample_size": 1024,
    "resolution": 1024,
    "resolution_type": "pixel_area",
    "cache_dir_vae": "cache/vae/cc12m-webshart",
    "webshart": {
      "cache_dir": "cache/webshart/cc12m",
      "shard_cache_gb": 25,
      "parallel_downloads": 4
    }
  },
  {
    "id": "alt-embed-cache",
    "dataset_type": "text_embeds",
    "default": true,
    "type": "local",
    "cache_dir": "cache/text"
  }
]

Add "max_num_samples": 65536 to train on a fixed-size subset. See SimpleTuner's documentation/DATALOADER.md ("Webshart Datasets") for all options, including webshart_optimize_captions — unnecessary for this dataset since captions are already embedded, but useful for sidecar-caption datasets that have not been optimized yet.

Regenerating

The caption embedding was produced with webshart >= 0.5.2:

webshart optimize-captions \
  --source laion/conceptual-captions-12m-webdataset \
  --metadata webshart/conceptual-captions-12m-webdataset-metadata \
  --destination caption-metadata \
  --shard-cache-dir cache/shards \
  --push-to-hub webshart/conceptual-captions-12m-webdataset-metadata \
  --path-in-repo data