Datasets:
image_id stringlengths 21 47 | source stringclasses 3
values | image_bytes unknown | ext stringclasses 2
values | condition_label stringclasses 124
values | condition_label_detail stringlengths 64 106 | fst int64 1 6 ⌀ | fst_provenance stringclasses 1
value | mst int64 | license stringclasses 3
values | attribution stringclasses 3
values | source_url stringclasses 3
values | fever_embedding listlengths 1.15k 1.15k | fever_blob_id int64 1 39.4k | width int64 | height int64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
fitzpatrick17k:2042ae11fac08fba6e15177fbafeb8c4 | fitzpatrick17k | [
255,
216,
255,
224,
0,
16,
74,
70,
73,
70,
0,
1,
1,
1,
0,
96,
0,
96,
0,
0,
255,
219,
0,
67,
0,
8,
6,
6,
7,
6,
5,
8,
7,
7,
7,
9,
9,
8,
10,
12,
20,
13,
12,
11,
11,
12,
25,
18,
19,
15,
20,
29,
26,
31,
30,
29,
... | jpg | sarcoidosis | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 5 | dataset_label | null | CC BY-NC-SA 3.0 | Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset, CVPR 2021 (github.com/mattgroh/fitzpatrick17k) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [
0.01242828369140625,
0.00347137451171875,
-0.01215362548828125,
0.00992584228515625,
-0.01202392578125,
0.01287078857421875,
-0.033050537109375,
-0.03955078125,
-0.010040283203125,
-0.00824737548828125,
-0.022705078125,
0.0035381317138671875,
0.11358642578125,
-0.0208740234375,
-0.002176... | 1 | null | null |
fitzpatrick17k:b1d2d9cd5adbf8d00c245c55ed9b9a3c | fitzpatrick17k | "/9j/4SxjRXhpZgAASUkqAAgAAAAKAA8BAgAGAAAAhgAAABABAgAOAAAAjAAAABIBAwABAAAAAQAAABoBBQABAAAAmgAAABsBBQA(...TRUNCATED) | jpg | lichen simplex | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 2 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.01427459716796875,-0.0002372264862060547,0.019927978515625,0.004177093505859375,-0.00514221191406(...TRUNCATED) | 2 | null | null |
fitzpatrick17k:b9677d43bff719c0303324b815f45ce3 | fitzpatrick17k | "/9j/4QDkRXhpZgAASUkqAAgAAAAIABIBAwABAAAAAQAAABoBBQABAAAAbgAAABsBBQABAAAAdgAAACgBAwABAAAAAgAAADEBAgA(...TRUNCATED) | jpg | scleromyxedema | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 2 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.0234527587890625,0.0035419464111328125,0.01409149169921875,-0.01502227783203125,-0.01936340332031(...TRUNCATED) | 3 | null | null |
fitzpatrick17k:342a75ba972d30bbff02b16ca81c9242 | fitzpatrick17k | "/9j/4QDkRXhpZgAASUkqAAgAAAAIABIBAwABAAAAAQAAABoBBQABAAAAbgAAABsBBQABAAAAdgAAACgBAwABAAAAAgAAADEBAgA(...TRUNCATED) | jpg | mucinosis | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 2 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.0179290771484375,-0.0169830322265625,0.01435089111328125,-0.0120086669921875,-0.0428466796875,-0.(...TRUNCATED) | 4 | null | null |
fitzpatrick17k:fe1d319bf00127bf0cb755abddf2ee15 | fitzpatrick17k | "/9j/4AAQSkZJRgABAgEBLAEsAAD/7gAOQWRvYmUAZAAAAAAB/9sAhAAKBwcHCAcKCAgKDwoICg8SDQoKDRIUEBASEBAUEQwMDAw(...TRUNCATED) | jpg | striae | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 1 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.02423095703125,-0.030792236328125,0.0028533935546875,-0.0009179115295410156,-0.051544189453125,-0(...TRUNCATED) | 5 | null | null |
fitzpatrick17k:951559d1b2d52080c2bd4323a7bf58bf | fitzpatrick17k | "/9j/4QDkRXhpZgAASUkqAAgAAAAIABIBAwABAAAAAQAAABoBBQABAAAAbgAAABsBBQABAAAAdgAAACgBAwABAAAAAgAAADEBAgA(...TRUNCATED) | jpg | folliculitis | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 3 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.0191650390625,0.0034580230712890625,0.0130767822265625,-0.01580810546875,0.0037097930908203125,-0(...TRUNCATED) | 6 | null | null |
fitzpatrick17k:3c51c4031263c4fabc496dc1bcad04ff | fitzpatrick17k | "/9j/4AAQSkZJRgABAQEBLAEsAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RDgsLEBYQERM(...TRUNCATED) | jpg | keloid | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 4 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.033782958984375,-0.003173828125,0.038421630859375,0.020965576171875,-0.0015058517456054688,-0.033(...TRUNCATED) | 7 | null | null |
fitzpatrick17k:2753aae451d232c73026a18fa18d9a64 | fitzpatrick17k | "/9j/4AAQSkZJRgABAgEBLAEsAAD/7gAOQWRvYmUAZAAAAAAB/9sAhAAKBwcHCAcKCAgKDwoICg8SDQoKDRIUEBASEBAUEQwMDAw(...TRUNCATED) | jpg | striae | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 2 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.0246734619140625,-0.006145477294921875,-0.011993408203125,0.0009274482727050781,-0.05197143554687(...TRUNCATED) | 8 | null | null |
fitzpatrick17k:15dd1eb153ea63ea9a1638296520460d | fitzpatrick17k | "/9j/4QDmRXhpZgAASUkqAAgAAAAFABIBAwABAAAAAQAAADEBAgAcAAAASgAAADIBAgAUAAAAZgAAABMCAwABAAAAAQAAAGmHBAA(...TRUNCATED) | jpg | milia | nine_partition_label=benign epidermal; three_partition_label=benign | 2 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [-0.0091552734375,0.0030059814453125,0.0191802978515625,-0.01218414306640625,-0.0032596588134765625,(...TRUNCATED) | 9 | null | null |
fitzpatrick17k:87e33bc3d8bbeb9522da95e4d8459ce2 | fitzpatrick17k | "/9j/4QCwRXhpZgAASUkqAAgAAAAFABIBAwABAAAAAQAAADEBAgAcAAAASgAAADIBAgAUAAAAZgAAABMCAwABAAAAAQAAAGmHBAA(...TRUNCATED) | jpg | scleroderma | nine_partition_label=inflammatory; three_partition_label=non-neoplastic | 4 | dataset_label | null | CC BY-NC-SA 3.0 | "Groh et al., Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fit(...TRUNCATED) | https://github.com/mattgroh/fitzpatrick17k (metadata CSV); images via HF mirror huggingface.co/datasets/czty/SkinGPTX-Images (R12_FourDatasets_20260929/Fit17k), md5-verified against canonical CSV | [0.0139617919921875,-0.0031337738037109375,-0.02593994140625,-0.01045989990234375,-0.030868530273437(...TRUNCATED) | 10 | null | null |
SkinTone-540k: Tone-annotated dermatology images with FEVER embeddings
Built with FEVER Multimodal Database: AI for good, for real.
What is in this Dataset?
The largest redistributable skin-tone-annotated dermatology corpus: ~540k images, each with condition labels, per-image license lineage, 1152-d FEVER embeddings, a cross-dataset dedup graph, label-noise flags, and provably leakage-free train/val/test splits.
Why did we do this?
Lowdown Labs SPC agrees with Dr. Timnit Gebru - "AI needs to work for us."
Dermatology AI underperforms on dark skin, many datasets are imbalanced, poorly curated, and cause disparate impact. We built the FEVER Multimodal Database to rapidly curate large image and video datasets for supporting AI training. We figured this is a perfect first debut.
The public benchmark for this (DDI) has 656 images and a no-derivatives license; Fitzpatrick17k has noisy web labels. We unified every redistributable public source, verified tone labels against ITA colorimetry on 26.5k human-labeled images, estimated tone where it was missing, and used FEVER's rapid embeddings to track duplicates across sources and build splits with zero near-duplicate leakage between train and test. BOOM - big dataset!
Dataset composition
| Source | Images | Skin-tone labels | License |
|---|---|---|---|
| Fitzpatrick17k | ~17k | FST I–VI | CC BY-NC-SA 3.0 |
| SCIN | ~10k | FST + Monk MST | SCIN Data Use License |
| PAD-UFES-20 | ~2300 | FST | CC BY 4.0 |
| HAM10000 | ~10k | ITA-estimated | CC BY-NC 4.0 |
| ISIC Archive | ~500k | ITA-estimated | per-image CC0/CC-BY/CC-BY-NC |
Column names
image_id, source, image_bytes, condition_label(+ detail), fst, fst_provenance, mst, ita (estimated + degrees), ita_band, mst_knn (embedding-propagated Monk tone), label_confidence, suspect_label, split, license, attribution, source_url, fever_embedding, width, height
Accuracy validation
- FST↔ITA consistency on 26.5k human-labeled images, Spearman ρ, per-band medians reported
- Label-noise audit, kNN condition agreement; suspect-label flags per image
- Dedup audit, near-duplicate graph across sources
- (SCIN's 15 documented dupes found as positive control)
- Leakage-free splits, including no cross-split near-dupes, curated and collated with FEVER
How we did it
Every image was embedded with the FEVER Multimodal Database @ 224px resolution and we measured elapsed wall-clock:
- On CPU: ~60k images end-to-end in ~90 min on a c7i.8xlarge - 10.6 img/s sustained (including ITA etc)
- On GPU: full ~540k corpus on a single NVIDIA L40S (g6e.xlarge)
Ethical usage
Research use. Not a diagnostic device or for direct diagnostic usage. Non-commercial release honoring the original NC sources; per-image license and attribution given for every row. Estimated tone labels are marked estimated; set your own thresholds via the raw ITA values. Web-sourced images carry known label noise which we attempt to quantify. Use our embeddings as you wish to cluster, explore, or understand the dataset.
License
CC BY-NC 4.0 for the compilation; per-image licenses vary by source.
SCIN rows are under the SCIN Data Use License (attribution required).
Fitzpatrick17k rows are CC BY-NC-SA 3.0 (share-alike applies to those rows).
Explore it
Try the SkinTone-540k Explorer - embedding similarity search over all 542,504 images, the cross-dataset dedup cluster browser, and distribution dashboards, all running on the precomputed FEVER embeddings in this repo.
- Downloads last month
- -