SMAT: Simple and Efficient Merge-Aware Training
Paper • 2609.33437 • Published • 5
Datasets:
Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
SMAT: tiny notebook data
Fixed, balanced subsets for the SMAT demonstration. Code and notebook. These are illustrative demo subsets, not the paper's benchmark splits.
Two tasks: CIFAR-10 object recognition and SVHN cropped digit recognition. Each task has 2,000 training, 500 development and 2,000 test images. The notebook uses the frozen recipe and does not search on development data.
train.npz, dev.npz, and test.npz contain cifar10_images, cifar10_labels,
svhn_images, and svhn_labels. Images are original 32×32 RGB uint8 arrays;
labels are int64, 0–9. Load with numpy.load(..., allow_pickle=False).
Sources: CIFAR-10 revision
0b2714987fa478483af9968de7c934580d0bb9a2 and
SVHN revision
f9e1717d73324ebbc1ece84267ccd80d0aca0690 (cropped_digits).
Source licenses and terms apply; no new license is granted for the images.
CIFAR-10's upstream card lists its license as unknown; SVHN's original
terms restrict use to non-commercial purposes.
manifest.json records source hashes and exact source-row indices. Training
and development come from the official training splits; test comes from the
official test splits. Selected subsets have no identical decoded images
across splits. The split procedure is in prepare_subsets.py.
- Downloads last month
- 37