Modelpublic

aarontseng/bert-base-multilingual-cased-awesome-align-finetrans-zh

sourceHugging Facemitupdated 14d agoView on Hugging Face

0likes23downloads

Model Card

aarontseng/bert-base-multilingual-cased-awesome-align-finetrans-zh

Fine-tuned `google-bert/bert-base-multilingual-cased` with awesome-align master (TLM + SO, no CO) on FineTranslations Chinese (cmn_Hani) plus TsinghuaAligner non-test sure gold (v1 test / awesome-align examples/zhen excluded). Tsinghua non-test sure gold n=28610.

Setting	Value
Objectives	TLM + SO + sure gold SFT
Align layer	8
Extraction	softmax intersection @ 0.001
Checkpoint	`checkpoint-9500`
LR / batch	2e-05 / 8x4
Languages	zh-en only

Eval is awesome-align examples/zhen (same files as the README Zh-En row). Selected by precision, not AER.

pair	P	R	AER %
zhen	0.8732	0.8641	13.13

Awesome README softmax AER (lower better): Zh-En 18.1 (w/o FT) / 13.4 (multilingual FT, no CO).

Use layer 8 hidden states for awesome-align softmax alignment.