MNIST-derived digit samples converted into 840-token text sequences using a 64-character quantization scheme. Each entry includes dual 'up' and 'down' versions of the digit to facilitate rotation-invariant feature learning in Transformer-VAE models.
Use Cases
- Test interpolation quality for Transformer-VAEs using the text sequence data
- Train a digit classifier using the label column and the quantized text representations
- Analyze the impact of BPE tokenization on structured image-to-text data
- Convert generated text back into image arrays using the .text_to_array() utility for visual verification
Strengths
- 840-character sequences representing a 30x28 grid of quantized pixel values
- Pixel intensities mapped to a specific 64-character vocabulary
- Dual-orientation samples labeled with 'up' and 'down' prefixes to encourage rotation invariance
- Removal of whitespace to maximize Byte Pair Encoding (BPE) compression efficiency