workfunction/McBopomofoLM-models

McBopomofoLM models (Core ML, Apple Neural Engine)

Core ML conversions of the SlothE models used by McBopomofoLM (branch lm), a fork of McBopomofo that uses the models to improve candidate selection and to suggest corrections for mistyped syllables. Both models run on the Apple Neural Engine only (computeUnits = CPU_AND_NE).

Revisions: main and tag v3.1.2 match McBopomofoLM v3.1.x. Tag v2.1.1 is the runtime for McBopomofoLM v2.1.1, which has no next-token functions.

Contents

Path	What
runtime/	The exact runtime bundle shipped in McBopomofoLM v3.1.2 (Contents/Resources/SlothE): compiled .mlmodelc, embedding table, vocabularies, the next-token first-char table dec_next_first.tsv, and runtime-manifest.txt. The app checks the size and SHA-256 of each file against the manifest before it loads that file.
mlpackage/enc25m_multi_pal2_emb.mlpackage	SlothE-T 25M encoder, 2-bit palettized. The ternary weights make this lossless. Multifunction package with functions L8/L16/L32/L64/L256. The embedding lookup runs in the caller, from runtime/enc25m_embed_f16.bin.
mlpackage/dec_mf_tn_fp16.mlpackage	SlothE decoder pred_q35_60m (Qwen3.5 architecture), fp16, one weight file shared by two function families. t16/t32/t64/t96 (batch 3) return per-token log-probabilities of given texts; they are bit-identical to v2.1.1's decoder. n16/n32/n64/n96 (batch 1; input ids right-padded and last = index of the last real token) return the full next-token log-softmax over the 16k vocabulary. Gated DeltaNet runs in fixed 16-token chunks.
scripts/	Conversion and bundling scripts, research-grade. They import helpers from the author's research workspace, which is not included, so they document the conversion and do not run on their own. MCBPMF_LM_WORK sets the workspace root. build_next.py builds the n* functions; build_dec_tn.py merges them with the t* functions.

Measured on an M2 (ANE):

• encoder: 100% of ops on the ANE, about 0.8 ms per call;
• decoder t*: 86–95% on the ANE, 2.6–7.9 ms per call depending on length;
• decoder n*: 83–94% on the ANE, 2.3–3.1 ms per call.

Against HF fp32 on CPU, the n* functions agree on the top-1 token for 95.5% of 288 prefixes; every mismatch is a near-tie. The largest log-prob difference over the top-50 tokens is 0.31 nats at the median and 0.73 at P95, from fp16 on the ANE. The first compile of all functions takes about 3 minutes.

Sources and licenses

• Weights: Luigi/sloth-ime-models, Apache-2.0.
• char2id.tsv: converted from enc/char2id.json in Luigi/slothing-web, Apache-2.0.
• Variant classes: derived from OpenCC TWVariants.txt and HKVariants.txt, Apache-2.0.
• Conversion: Apache-2.0 (LICENSE). See NOTICE.txt.