10 digit classes represented as 224-token text strings derived from 64-character pixel quantization of downsampled MNIST images. Each sample includes both 'up' and 'down' orientations to support the development of rotation-invariant features in generative models. The data is formatted for testing interpolation quality specifically within Transformer-VAE architectures.
Use Cases
- Evaluate latent space interpolation in Transformer-VAEs using the 224-token text sequences
- Train a classification model to predict the numeric label column from the quantized character strings
- Benchmark rotation-invariant feature extraction by comparing the 'up' and 'down' variants within the text data
Strengths
- 224 tokens per sample arranged in a 16x14 grid format within the text column
- Pixel intensities are quantized into a discrete 64-character vocabulary
- Every digit sample is provided in both 'up' and 'down' orientations to encourage rotation invariance
- Images are reduced to 1/4 of their original area via max pooling before text conversion