manga-ocr (base), 8-bit GGUF

manga-ocr by Maciej Budyś (kha-white), converted to GGUF with 8-bit weights. It reads Japanese text in manga speech bubbles and other short text: a ViT encoder with a BERT decoder.

Attribution

• Original work: kha-white/manga-ocr-base, revision aa6573bd10b0d446cbf622e29c3e084914df9741, © Maciej Budyś.
• Licence: Apache License 2.0, the same as the original.

Changes from the original

• The PyTorch checkpoint (pytorch_model.bin) is stored as GGUF, under the original tensor names.
• Two-dimensional .weight matrices of linear layers are quantised to q8_0 (blocks of 32, one scale per block). Where a row's length is not a multiple of 32 it is padded with zeros at the end.
• Embeddings and other matrices are stored as float16; norms and biases as float32.
• vocab.txt is the original tokenizer vocabulary, unchanged.
• No weights were retrained or fine-tuned.

Files

Accuracy

On 40 speech bubbles with hand-checked text, the 8-bit model reads exactly as well as the original float32 model (4.1% character error rate for both).