predictor v2.1 full weights (TW-chat register FT)

predictor v2.1 full weights (TW-chat register FT)

card: decoder v2.1 section (register FT, honest evals) + repro pointer

predictor v2.1: TW-chat register fine-tune — chat 10.9->18.3 top-1, fresh-C4 unchanged; drop-in

predictor v2 full weights (6.1M-line corpus)

predictor v2: retrained on 6.1M lines — honest fresh-corpus 34/46 vs old 4.6/12 (leak-inflated 47.3/75.8); drop-in

retitle as the Sloth IME model-family repo (renamed from slothe-t-25m-zhuyin)

plot: all-measured (homophone-hard vs on-device latency, 4 models, projection gone)

decoder full weights: 60M dense-Qwen3.5 GDN next-word predictor (HF format, fp32 safetensors + tokenizer)

Files: list 12M GGUF (shipping default), 12M fp32 master, 60M predictor

12M weights: 12m/roles.json

12M weights: 12m/slothe_config.json

12M weights: 12m/model.safetensors

add 60M Q4 next-word predictor (qwen35 GDN hybrid, 47.3/75.8 held-out, 8.5ms/word on BOOX @4t )

add 12M ternary 256x12 sibling (9.65MB TQ2_0, 84%/84%, 9.3ms@4t on BOOX)

latency correction: 25M ~9ms was projected; measured 18.5ms@4t/31.9ms@2t; add measured 12M sibling

Render score-vs-latency chart as PNG for reliable HF display

Add score-vs-latency chart; fix pipeline_tag to token-classification

SlothE-T 25M ternary Zhuyin→zh-TW IME model: ternary GGUF (TQ2_0) + fp32 master + reproducibility toolchain

initial commit