notesum.ai

Published at November 11

Stronger Models are NOT Stronger Teachers for Instruction Tuning

cs.AI

cs.CL

Released Date: November 11, 2024

Authors: Zhangchen Xu¹, Fengqing Jiang¹, Luyao Niu¹, Bill Yuchen Lin², Radha Poovendran¹

Aff.: ¹University of Washington; ²Allen Institute for AI

Arxiv: http://arxiv.org/abs/2411.07133v2

Refer to caption

Response	AlpacaEval 2		Arena-Hard	AP
Generator Model	LC (%)	WR (%)	WR (%)	(%)
Gemma-2-9b-it	16.09	13.70	13.7	14.90
Gemma-2-27b-it	13.93	13.31	12.4	13.17
Llama-3-70b-Instruct	10.55	10.68	6.7	8.62
Llama-3.1-70b-Instruct	9.52	10.10	8.3	8.91
Qwen2.5-7B-Instruct	13.50	14.33	10.6	12.05
Qwen2.5-72B-Instruct	19.20	21.01	13.1	16.15
GPT-4	6.63	5.70	4.8	5.72