Single-file ONNX of the encoder-only CATT model (Abjad AI, Apache-2.0). CATT-EO is non-autoregressive — a transformer encoder followed by a linear classifier head — so TigreGotico stitched the two stages into one ONNX with onnx.compose. The stitched graph is byte-identical (max|Δ| = 0.0) to the original two-stage encoder→head pipeline.

I/O. Inputs src (int64, [batch, seq] Buckwalter ids) and src_mask (bool, [batch, 1, seq, seq]). Output [batch, seq, 18] tag logits. Tokenizer: tashkeel_tokenizer_onnx.py (+ bw2ar.py, utils.py) — Buckwalter encoding and the 18-tag tashkeel scheme. See catt_models_onnx.py for an end-to-end example.

Benchmark. ~4.27 % DER on the broad TigreGotico/arabic_diacritized_text test (much lower on CATT's own narrow benchmark — distribution-dependent). The int8 file is dynamic-quantized; expect small divergence from fp32.

Part of the Arabic Diacritizers / Tashkeel collection. Original CATT © Abjad AI (Apache-2.0); this is a stitched re-export with attribution.

catt_ed_encoder.onnx + catt_ed_decoder.onnx (+ .int8) are the autoregressive encoder-decoder model — the decoder runs once per output position (causal mask), so it is two ONNX (not stitchable) and much slower than CATT-EO. In text2tashkeel it is the catt-ed / catt-ed-int8 model.