tesseract-frk-GGUF

Tesseract German Fraktur — GGUF

Native CrispEmbed GGUF port of the Apache-2.0 Tesseract frk LSTM language model. It recognizes German Fraktur line images, including long-s (ſ) and historical characters in the upstream output alphabet.

Source: tesseract-ocr/tessdata frk.traineddata

• Source SHA-256: 7cd1b541e9d3884b9546a7d292c4f349cdb45ff646b4b43166dfc5099a8ad1a1
• VGSL: [1,48,0,1Ct3,3,16Mp3,3Lfys64Lfx96Lrx96Lfx384O1c1]
• Input height: 48 pixels
• Output alphabet: 100 Unicode tokens / 99 CTC output codes
• Parameters: 933,763
• Architecture: convolution + max-pool + four recurrent LSTM stages + CTC projection

Files

The Q8 variant keeps output.weight and output.bias at F32 because they are the sensitive CTC decision boundary. Recurrent matrices are Q8_0; the input convolution and smaller matrices remain F32. This is the same precision-first policy used for other CrispEmbed OCR GGUFs.

Usage

Pass a single-line grayscale crop to the native Tesseract LSTM path:

crispembed -m tesseract-frk-q8_0.gguf --ocr line.png

For full pages, run a text detector/line segmenter first and recognize each deskewed crop independently. This model is a line recognizer, not a page layout or text-detection model.

Validation

The native implementation was compared with a pure-Python reference generated from the same frk.traineddata and the same 200×80 grayscale crop:

The reference activation dump is retained outside Git at /Volumes/backups/ai/crispembed-gguf/tesseract-frk-ref-line.gguf.

License

Apache-2.0, following the upstream Tesseract language data. Preserve the upstream source URL, checksum, and license when redistributing this GGUF.

Provenance and EU AI Act Art. 53 note

• Upstream model: tesseract-ocr/tessdata.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.