aloudreader/uk-stress-crossencoder-ct2-int8

uk-stress-crossencoder, CTranslate2 INT8

aloudreader/uk-stress-crossencoder (XLM-R cross-encoder that picks the sense, and so the stress, of a Ukrainian homograph from its gloss) converted to CTranslate2 with INT8 weights.

	PyTorch FP32	CT2 INT8
size	1,061 MB	267 MB
CPU latency per target	100 ms	33 ms

Compared with the FP32 model on 1,226 targets from held-out silver sentences: the same top reading in 97.9% of them; the margin between the top two readings tracks FP32 with r = 0.984 (FP32 = 0.99 x INT8 + 0.09), so the served threshold 0.5 is kept. serving_manifest.json (the forms, their candidate senses and glosses, the threshold) is the original model's.

The CT2 encoder returns hidden states; the classification head is in classifier_head.npz (dense_w, dense_b, out_w, out_b: tanh(h W1 + b1) W2 + b2 on the first token).

Used by the light mode of the accentuk package (with spaCy uk_core_news_sm for morphology). Measured on 2026-10-13, word accuracy:

test	service / full mode	light mode
200 most frequent ambiguous forms of Common Voice uk, stress heard in the recordings	95.88%	95.54%
modern-text gold, 3,887 words	98.07%	97.63%
lang-uk stress benchmark, words / heteronyms	96.10% / 84.62%	96.10% / 84.62%