ngqtrung/aidm-dogs-vs-cats-data

aidm-dogs-vs-cats-data

The images of a dogs-vs-cats image-classification study, in the directory layout that results/splits.json indexes. The run registry, the splits file, the report tables and figures live in the companion results repo; the checkpoints live in the companion model repo.

Provenance

These images are the Kaggle Dogs vs. Cats competition data (<https://www.kaggle.com/c/dogs-vs-cats>). They are not original work and they are not relicensed here. They are redistributed only so that the splits and the published results are reproducible from one place: the index lists in results/splits.json are positions into this exact file layout, and without the files those indices name nothing. The original Kaggle competition rules and data terms apply to every file in this repo. If you need the data for anything other than reproducing these results, take it from Kaggle.

Generated by scripts/90_publish_hf.py on 2026-09-22 08:17 UTC. Every count, fingerprint and metric below was read from the run registry during that run, so this card describes the runs that exist, not the runs that once existed.

Splits

Every count below is read out of results/splits.json at publish time, so this card cannot describe a split the file does not contain.

Training subsets

The sample-efficiency runs train on nested subsets of train, indexed by the same file. The largest subset is the whole training split.

Cross-validation folds

5 folds over train + val_dev, each an (train indices, validation indices) pair.

Fingerprint

splits_fingerprint = 0c31c0f203df (sha1 over the canonicalised contents of results/splits.json, first 12 hex digits). Every run record in the companion results repo carries this value, so a result can be traced back to the exact index lists that produced it.

Duplicates

The dataset audit found and removed 19 duplicate image(s) before the splits were drawn (meta.duplicates_removed). A duplicate that straddles the train and validation splits leaks the label, so the removal happens before the split, not after.

The splits were drawn with seed 1337.

Files

• —datasets/{train,val,test}/<class>/* -- the image files themselves, in the exact directory layout the index in results/splits.json is built over. (25500 file(s), 1.1 GiB)
• —README.md -- this card, generated from the registry at publish time.