transcoda-omr-GGUF

<!-- crispembed-license-notice -->

License: CC BY-NC 4.0 — non-commercial use only. Transcoda's encoder is fine-tuned from facebook/convnextv2-tiny-22k-224 (per-stage weight cosines 0.95-0.99, so not trained from scratch), and Meta licenses the ConvNeXt V2 ImageNet pre-trained weights under CC BY-NC (see the ConvNeXt V2 README). The upstream cc-by-4.0 tag therefore cannot grant commercial use. An earlier version of this card said cc-by-4.0; that was incorrect. Evidence: license audit.

<!-- /crispembed-license-notice -->

Transcoda-59M zero-shot OMR — GGUF (CrispEmbed)

GGUF conversions of btrkeks/transcoda-59M-zeroshot-v1 for the dependency-free C/C++ CrispEmbed inference engine (ggml; CPU / Metal / CUDA / Vulkan).

End-to-end zero-shot Optical Music Recognition: a full-page score image → **kern (Humdrum) symbolic transcription. ConvNeXt-V2-Tiny encoder + 8-layer RoPE cross-attention Transformer decoder, 58.8 M parameters. OMR-NED SOTA on real historical scans.

Files

Usage

# auto-download via the CrispEmbed model manager
crispembed --ocr page.png -M transcoda

# or explicit
crispembed -m transcoda-q8_0.gguf --ocr page.png

The engine reproduces the reference preprocessing (resize to width 1050 preserving aspect, top-crop / white-pad to height 1485, normalize to [-1, 1]) and greedy decode (repetition_penalty=1.1) internally. Grammar-constrained / beam decoding from the original project are not (yet) part of this engine.

Provenance & validation

Converted with models/convert-transcoda-to-gguf.py. The CrispEmbed engine (src/transcoda_ocr.cpp) is a clean-room implementation written from the paper, the model's config/data files, and a reference-activation oracle — it does not copy the AGPL reference code. Per-stage cosine vs the reference is 1.000000 at F32 (all encoder + decoder stages, CPU and Metal); the greedy **kern output is byte-identical to the reference on the validation page.

License & attribution

These weights are redistributed under CC-BY-4.0, inherited from the source model. Attribution (required by CC-BY-4.0):

• Model: btrkeks/transcoda-59M-zeroshot-v1 by btrkeks — licensed CC-BY-4.0.
• Paper: Transcoda, arXiv:2605.10835.
• Project code: https://github.com/btrkeks/transcoda (AGPL-3.0; not used to produce this engine — see clean-room note above).
• Encoder backbone: facebook/convnextv2-tiny-22k-224.

If you use this model, please cite the original Transcoda work.

Provenance and EU AI Act Art. 53 note

• Upstream model: btrkeks/transcoda-59M-zeroshot-v1 — published by btrkeks.
• Upstream licence: cc-by-4.0. 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.