840-token textual representations of MNIST digits quantized into a 64-character vocabulary. Each sample includes both 'up' and 'down' orientations to facilitate the training of Transformer-VAEs on rotation-invariant features.
Use Cases
- Evaluate latent space interpolation quality in Transformer-VAEs using the 840-token text sequences
- Train generative models to reconstruct digits by predicting the 64-character quantized tokens in the text field
- Benchmark rotation-invariant feature extraction by comparing the 'up' and 'down' versions of the same digit
Strengths
- Quantizes grayscale pixel values into 64 distinct character tokens
- Each sample contains 840 tokens structured in a 30x28 grid format
- Includes dual 'up' and 'down' versions for every digit to support rotation-invariant learning
- Provides structured row indexing from 00 to 29 within the text column