workfunction's picture
McBopomofoLM v2.1.1 Core ML models (SlothE-T 25M encoder, pred_q35_60m decoder)
ff5f59d
Raw History Blame Contribute Delete
809 Bytes
McBopomofoLM v2 bundles third-party parts (side-by-side build next to stock McBopomofo):
- SlothE-T 25M encoder (enc25m.mlmodelc, enc25m_embed_f16.bin; converted to Core ML from the published weights) and its syl_vocab / syl2legal tables: huggingface.co/Luigi/sloth-ime-models, Apache-2.0.
- SlothE decoder pred_q35_60m (dec60m.mlmodelc, dec_tokenizer.json; converted to Core ML from the published weights): huggingface.co/Luigi/sloth-ime-models, Apache-2.0.
- char2id table (char2id.tsv, converted from enc/char2id.json): huggingface.co/spaces/Luigi/slothing-web, Apache-2.0.
- Orthographic variant classes derived from OpenCC TWVariants.txt / HKVariants.txt, Apache-2.0 (OpenCC-LICENSE.txt).
v2 runs the models with Core ML on the Apple Neural Engine only. It contains no ggml, llama.cpp or libslothe code.