glyph-models

Glyph decoder resources

Core ML models and tables for Glyph, an open-source iPhone swipe keyboard. SwipeAREncoder / SwipeARStep: the autoregressive swipe decoder (research run ar_mixed_s1, TCN trunk + 2-layer transformer letter decoder, 1.7M parameters, fp32); ar_weights.bin: the same tensors raw, for the keyboard's Accelerate first pass. LM_smollm2_135m_fp16: SmolLM2-135M fp16 gather model, the default sentence LM of the fused sentence search (research #115/#116); SwipeLM: distilgpt2 fp16, the previous default and the fallback. lexicon.bin: 301,508-word trie (FUTO training words blended with wordfreq English) with unigram log-probabilities. ilm.bin: the encoder's internal LM per trie node; priors_smollm2_135m.bin / priors.bin: each LM's marginal prior per trie node, priors_surfaces_smollm2_135m.bin the same for the apostrophe spellings. gpt2/: distilgpt2's tokenizer files (SmolLM2's are in git). bench_gestures.json: the replay-benchmark gestures. *goldens.json: test vectors the Swift port is checked against. Regenerated by the tools/export_*.py scripts in the repo; how each was chosen is in research/README.md.