notesum.ai

Published at December 9

JAPAGEN: Efficient Few/Zero-shot Learning via Japanese Training Dataset Generation with LLM

cs.CL

Released Date: December 9, 2024

Authors: Takuro Fujii¹, Satoru Katsumata²

Aff.: ¹Yokohama National University; ²Retrieva, Inc.

Arxiv: http://arxiv.org/pdf/2412.06738v1

Refer to caption

Diversity (%)
Dataset	MAR.	JSTS*	JNLI	JCoLA
Gold	40.53	72.93	72.94	56.66
Zero-shot	91.67	74.89	69.97	65.80
w/ KADG	84.97	76.12	73.13	78.91
Few-shot	90.25	81.80	78.28	67.15
Label Correctness (%)
Gold	99.06	0.137	98.01	96.28
Zero-shot	99.97	1.540	35.11	66.34
w/ KADG	99.96	1.540	39.37	63.94
Few-shot	99.90	1.094	50.16	63.33