README.md

---
license: cc0-1.0
---

This dataset is a shuffled list of downloadable CC0 image titles and URLs from [Smithsonian Open Access](https://github.com/Smithsonian/OpenAccess).
Some images may be omitted due to limitations or oversights in the preprocessing pipeline, but there's no deliberate curation.

This dataset only contains metadata; a tool like https://github.com/rom1504/img2dataset can be used to download the actual images:

```bash
img2dataset --url_list data --output_folder data_files \
   --input_format "parquet" --output_format files \
   --caption_col "text" --url_col "url" \
   --image_size 256 --min_image_size 128 --resize_mode keep_ratio --resize_only_if_bigger true \
   --processes_count 16 --thread_count 128
```

This dataset includes human-authored titles and metadata directly from the SOA database.

AI Picasso have also created [`soa-full-florence2`](https://huggingface.co/datasets/aipicasso/soa-full-florence2) containing synthetic captions made with Florence-2.

![image/png](https://cdn-uploads.huggingface.co/production/uploads/630447d40547362a22a969a2/wGhISvXOIwNsHCy3wluD2.png)

The `soa-full` dataset is dominated by botanical images from NMNH. Here's what a random selection looks like with all NMNH images excluded:

![](random_non_nmnh.png)

vs. a true-random selection:

![](true_random.png)