cyttic/hebrew-ocr-android

Hebrew handwriting OCR: on-device model files

Files downloaded by the Android app (github.com/cyttic/android-hwrg), which runs the whole pipeline on the phone.

File	What
line_int8.onnx	CTC line model cyttic/exp35-trocr-base-en-ctc-seqhead-allreal-synpunct (TrOCR-base ViT encoder + 2-layer transformer CTC head, 198 classes), ONNX opset 17, dynamic int8 (all constant-weight MatMuls except the ViT layer-2 MLP fc2, kept in fp32). Inputs: pixel_values (B, 3, 384, 384) in [-1, 1], cols (B,) int64; output log_probs (B, 288, 198).
line_large_int8.onnx	CTC line model cyttic/exp35-trocr-large-en-ctc-seqhead-allreal-synpunct (TrOCR-large ViT encoder, 24 layers + 2-layer transformer CTC head), same inputs/outputs; dynamic int8 except the ViT layer-0 MLP fc2 (fp32). Decoded with the punctuation-aware LM fusion (alpha 0.3, beta 0.6, LM on Hebrew letters + space).
charset.json	chars[k] = character of class k; chars[0] = CTC blank.
heb_char6_kn.bin.gz	Character 6-gram LM, interpolated Kneser-Ney (Ben-Yehuda + DictaGen text), binary layout in tools/export_lm.py.
mlm_tiny_int8.onnx	Language check (optional second pass): dicta-il/dictabert-tiny by Dicta (CC BY 4.0, not MIT), converted to ONNX; the 128k x 312 word table in int8 with one scale per row, everything else fp32. Inputs input_ids, attention_mask (R, L) int64 with [MASK] at pos (R,), tgt (R,) = the original token; output logp (R,) = ln P(tgt at pos). The app keeps the line decoder's 16 best lines and picks the one with the best beam score + 0.25 x PLL + 2.0 x word pieces. Made by tools/export_mlm.py.
dictabert_vocab.txt	dictabert-tiny's WordPiece vocabulary (line number = token id; CC BY 4.0, Dicta).

The line finder is PaddlePaddle/PP-OCRv5_server_det_onnx (Apache-2.0), downloaded from its own repo.

License: MIT for our files; mlm_tiny_int8.onnx and dictabert_vocab.txt are derived from dicta-il/dictabert-tiny and stay under its CC BY 4.0 license (Shmidman et al., DictaBERT, 2023).