Datasets
Korfiatis, N., Avlonitis, M., Kourouthanassis, P. E., Karyotis, V., Ghinis, S., Gjoni, V., & Gkinis, S. (2026). G3 Corfu Butterflies: A Small-Scale Benchmark for Long-Tailed Multi-Label Attribute Recognition (Version 1.1) [Data set]. Harvard Dataverse. https://doi.org/10.7910/DVN/W6UMIR
@dataset{korfiatis2026g3,
title = {G3 Corfu Butterflies: A Small-Scale Benchmark for Long-Tailed Multi-Label Attribute Recognition},
author = {Nikolaos Korfiatis and Markos Avlonitis and Panos E. Kourouthanassis and Vasileios Karyotis and Stamatis Ghinis and Vojsava Gjoni and Spiros Gkinis},
url = {https://doi.org/10.7910/DVN/W6UMIR},
doi = {10.7910/DVN/W6UMIR},
year = {2026},
date = {2026-01-01},
publisher = {Harvard Dataverse},
edition = {1.1},
abstract = {G3 Corfu Butterflies is a public benchmark for small-scale, long-tailed, multi-label attribute recognition. It is designed as a tight, reproducible test-bed for computer-vision methods that target data-efficient learning, multi-label classification with strong class imbalance, and compositional attribute transfer. The subject matter (butterfly photographs collected in Corfu, Greece, by Stamatis Ghinis, Vojsava Gjoni, and Spiros Gkinis) is the data substrate; the contribution is the benchmark layer built on top of it. Contents of v1.0: - 165 high-resolution JPEG images (~73 MB total); - image-level multi-label annotations across three attribute families (colour, pattern, additional visual features), tokenised into a 45-entry vocabulary in a fixed order; - canonical evaluation splits: a locked 15% holdout test set (28 images) and five iteratively-stratified cross-validation folds over the remaining 137 images, all at seed 42; - two reference baselines with fully documented hyper-parameters: a frozen CLIP ViT-B/32 linear probe and a fine-tuned ViT-B/16; - two precomputed image-embedding matrices aligned row-for-row with metadata.csv (CLIP ViT-B/32, 512-d L2-normalised; torchvision ViT-B/16 ImageNet1k, 768-d pooled CLS) for downstream retrieval / clustering / weak-supervision studies; and - a data-descriptor paper (main.pdf) with full reproduction instructions, a descriptive-statistics section, and a Colab/Kaggle quickstart notebook. Reference numbers on the shipped splits: - CLIP linear probe: 5-fold macro-F1 0.291 ± 0.046; holdout 0.256. - Fine-tuned ViT-B/16: 5-fold macro-F1 0.395 ± 0.040; holdout ensemble macro-F1 0.342, micro-F1 0.563, mAP 0.443. Evaluation-active labels: because the benchmark is deliberately long-tailed and the holdout is small, 24 of 45 labels have at least one positive example in the locked holdout. Macro-averaged metrics are reported over those 24 labels; the remaining labels contribute only to the micro-averaged counts. This is fully documented in Section 3 of the paper. Primary uses (computer-vision research): data-efficient learning on O(10^2)-image budgets; long-tail multi-label classification and loss-reweighting studies; representation comparison of frozen foundation features against full fine-tuning; label-correlation modelling on a corpus small enough to iterate on; and compositional transfer across attribute families. Not included in v1.0: species labels, region-level masks/bounding boxes, GPS coordinates, and per-image timestamps. These extensions are listed under Future Work in the accompanying paper. Dedication: the dataset is dedicated to the memory of Professor Markos Avlonitis, whose sustained contributions to the computational study of biodiversity in the Ionian Islands shaped both the scientific culture in which this work was carried out and the motivation to build open, reusable resources of this kind. Please cite the dataset when using it in academic or applied research, and contact the corresponding author for collaboration or feedback.},
note = {Code: https://github.com/nkorf/g3-corfu-butterflies},
keywords = {},
pubstate = {published},
tppubtype = {dataset}
}
G3 Corfu Butterflies is a public benchmark for small-scale, long-tailed, multi-label attribute recognition. It is designed as a tight, reproducible test-bed for computer-vision methods that target data-efficient learning, multi-label classification with strong class imbalance, and compositional attribute transfer. The subject matter (butterfly photographs collected in Corfu, Greece, by Stamatis Ghinis, Vojsava Gjoni, and Spiros Gkinis) is the data substrate; the contribution is the benchmark layer built on top of it. Contents of v1.0: - 165 high-resolution JPEG images (~73 MB total); - image-level multi-label annotations across three attribute families (colour, pattern, additional visual features), tokenised into a 45-entry vocabulary in a fixed order; - canonical evaluation splits: a locked 15% holdout test set (28 images) and five iteratively-stratified cross-validation folds over the remaining 137 images, all at seed 42; - two reference baselines with fully documented hyper-parameters: a frozen CLIP ViT-B/32 linear probe and a fine-tuned ViT-B/16; - two precomputed image-embedding matrices aligned row-for-row with metadata.csv (CLIP ViT-B/32, 512-d L2-normalised; torchvision ViT-B/16 ImageNet1k, 768-d pooled CLS) for downstream retrieval / clustering / weak-supervision studies; and - a data-descriptor paper (main.pdf) with full reproduction instructions, a descriptive-statistics section, and a Colab/Kaggle quickstart notebook. Reference numbers on the shipped splits: - CLIP linear probe: 5-fold macro-F1 0.291 ± 0.046; holdout 0.256. - Fine-tuned ViT-B/16: 5-fold macro-F1 0.395 ± 0.040; holdout ensemble macro-F1 0.342, micro-F1 0.563, mAP 0.443. Evaluation-active labels: because the benchmark is deliberately long-tailed and the holdout is small, 24 of 45 labels have at least one positive example in the locked holdout. Macro-averaged metrics are reported over those 24 labels; the remaining labels contribute only to the micro-averaged counts. This is fully documented in Section 3 of the paper. Primary uses (computer-vision research): data-efficient learning on O(10^2)-image budgets; long-tail multi-label classification and loss-reweighting studies; representation comparison of frozen foundation features against full fine-tuning; label-correlation modelling on a corpus small enough to iterate on; and compositional transfer across attribute families. Not included in v1.0: species labels, region-level masks/bounding boxes, GPS coordinates, and per-image timestamps. These extensions are listed under Future Work in the accompanying paper. Dedication: the dataset is dedicated to the memory of Professor Markos Avlonitis, whose sustained contributions to the computational study of biodiversity in the Ionian Islands shaped both the scientific culture in which this work was carried out and the motivation to build open, reusable resources of this kind. Please cite the dataset when using it in academic or applied research, and contact the corresponding author for collaboration or feedback.
Varthis, E. (2020). Dorotheos Scholarios's Synoptic Index for Migne's Patrologia Graeca (Version 1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.4301071
@dataset{varthis2020dorotheos,
title = {Dorotheos Scholarios's Synoptic Index for Migne's Patrologia Graeca},
author = {Evagelos Varthis},
url = {https://doi.org/10.5281/zenodo.4301071},
doi = {10.5281/zenodo.4301071},
year = {2020},
date = {2020-01-01},
publisher = {Zenodo},
edition = {1},
abstract = {This data set contains the transcribed text of Dorotheos Scolarios Synoptic Index for Migne's Patrologia Graeca. The dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.},
note = {Web portal using the dataset: http://scholarios.graeca.org},
keywords = {},
pubstate = {published},
tppubtype = {dataset}
}
This data set contains the transcribed text of Dorotheos Scolarios Synoptic Index for Migne's Patrologia Graeca. The dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.