Jiamian/noisy-ostracods

Noisy Ostracods

Noisy Ostracods is a fine-grained, highly imbalanced microscopy image dataset for studying real-world label noise, robust image classification, and label correction. Images are cropped photographs of ostracod specimens collected from Hong Kong sediment samples. The original annotations were produced from expert taxonomic records and contain feature errors, label errors, negative examples, pseudo-classes, and open-set noise.

The release candidate contains the original image bytes in self-contained Parquet shards. No lossy format conversion or resizing is applied.

Dataset structure

Split	File records	Genus labels	Species labels
train	66,353	66,352	53,038
validation	7,117	7,117	5,465
test	7,113	7,113	5,458
total	80,583	80,582	63,961

The class index spaces contain 79 genus-level entries and 139 species-level entries, including negative classes. Missing genus/species fields are null; they are not encoded as a new class.

Each record includes:

• record_id: stable identifier derived from the legacy relative path.
• image: original TIFF or JPEG bytes as a Hugging Face Image.
• legacy_path: path used by the original training manifests.
• content_sha256: integrity and exact-duplicate identifier.
• image format, dimensions, and byte size.
• original noisy genus and species IDs and names where available.
• expert genus labels where explicitly present in the clean manifests.
• Confident Learning retention flags.
• legacy transition-matrix subset IDs where applicable. These compact IDs are preserved as distributed and are not asserted to map one-to-one to the current genus guide.

The expert_genus_available field is the authoritative availability flag. Records absent from the expert-cleaned CSVs are left null; this release does not infer whether absence means unreviewed or excluded as a feature error.

Loading

from datasets import load_dataset

dataset = load_dataset("Jiamian/noisy-ostracods")
train = dataset["train"]
print(train[0])

For inspection without downloading the entire dataset:

from datasets import load_dataset

train = load_dataset(
    "Jiamian/noisy-ostracods",
    split="train",
    streaming=True,
)
sample = next(iter(train))

For genus classification, select records where noisy_genus_id is not null. For species classification, select records where noisy_species_id is not null. Pin a Hub revision for reproducible experiments.

Known issues

• See the documentation of the paper for more issue.

Intended use and limitations

The dataset is intended for research on fine-grained image classification, learning with noisy labels, class imbalance, label-error detection, and dataset curation. The original labels must not be represented as fully verified ground truth. Model outputs are not authoritative taxonomic identifications and should not substitute for expert review in ecological or paleontological work.

Citation

@inproceedings{hu2024noisyostracods,
  title={Noisy Ostracods: A Fine-Grained, Imbalanced Real-World Dataset for Benchmarking Robust Machine Learning and Label Correction Methods},
  author={Hu, Jiamian and Hong, Yuanyuan and Chen, Yihua and Wang, He and Yasuhara, Moriaki},
  booktitle={Advances in Neural Information Processing Systems},
  year={2024}
}

Paper: https://arxiv.org/abs/2412.02313

Source code and legacy documentation: https://github.com/H-Jamieu/Noisy_ostracods