eden-keyboard-corrector

E.D.E.N. — Encoder-Decoder Enhancement Network (Keyboard Corrector)

The on-device neural typo corrector that powers the AI Keyboard ability in Rytability. A 16.8M-parameter character-level seq2seq Transformer that reads what you meant — using the misspelled word, the words before it, and (as you keep typing) the words after it — and fixes fat-finger scrambles, misspellings, and phonetic/dyslexic spellings, entirely on your device. Nothing you type ever leaves your phone.

What's in this repo

The Rytability app downloads these on demand when the AI Keyboard ability is enabled, compiles the Core ML packages on-device, and hands them to the keyboard extension through the shared app group. They are not useful on their own without the app's beam search, selection, and correction pipeline.

Model details

• Architecture: char-level encoder-decoder Transformer — d_model 384, 8 heads, 5+5 layers, ff 1024, 32-token vocab (a–z, apostrophe, space, specials). Input left-context [SEP] word [SEP] right-context; output the corrected word. Bidirectional by construction: the right context lets the keyboard revise earlier corrections as more of the sentence exists.
• Training (v20 lineage): cloze/denoise language pretraining on 7.4M quality-filtered sentences (FineWeb-Edu + conversational + formal), then specialization on realistic keyboard errors — spatial fat-finger noise, transpositions, drops, doubles, phonetic/dyslexic rewrites, real human misspelling corpora, and context-flip confusable supervision.
• Evaluation: 75.3% exact-fix on a 93-case held-out set of real typing mistakes (never trained on), 98.4% clean-word precision, 89% top-5 oracle.
• Precision: fp16 weights, fp32-safe compute paths chosen by the app at load time (with a correctness oracle).

Version

manifest.json > version — currently 20. The app checks the manifest to know when a newer model is available.

© RytechLabs LLC. All rights reserved. These weights are distributed for use by the Rytability app.