Hebrew grapheme-to-phoneme (G2P) training project for converting unvocalized Hebrew text into IPA.

Model: thewh1teagle/renikud

ModernBERT-base encoder trained from scratch (22 layers, 768 hidden, RoPE, Flash Attention) → three linear heads (consonant, vowel, stress) per Hebrew letter.

Each Hebrew letter gets exactly one output slot predicting a (consonant, vowel, stress) triple. Uses a custom character-level tokenizer built for Hebrew.

See docs/ARCHITECTURE.md for full design details.

See docs/TRAINING.md for data preparation, training commands, upload/download, and hyperparameters.

./scripts/train_bench.sh outputs/g2p-classifier/checkpoint-5000

Export a checkpoint to ONNX:

./scripts/export_onnx.sh outputs/g2p-classifier/checkpoint-5000

Then use the Python package:

from renikud_onnx import G2P

g2p = G2P("model.onnx")
print(g2p.phonemize("שלום עולם"))
# → ʃalˈom ʔolˈam

Or the Rust crate — see renikud-rs/.