Skip to content

Latest commit

 

History

History
25 lines (11 loc) · 347 Bytes

File metadata and controls

25 lines (11 loc) · 347 Bytes

Introduction

Motivation:

  • model越大不是越好,与人类意图不对齐(unalign)

Method:

  • fine-tuning with human feedback (RLHF)

Train process:

image-20230515213117585

Result:

  • 1.3B+RLHF>175B 性能不减,输出更真实

Discussion

  • RLHF是一个方向,用于对齐人类意图