Translation in progress Some entries still include text in the original language.
Failure-First·Indexed by this site · Originally published
论文:Chung 等把指令微调扩展至 1,836 个任务,训出 Flan-PaLM 与 Flan-T5
[Daily Paper] Scaling Instruction-Finetuned Language Models
AI summary
Chung 等人的论文把指令微调扩展到 1,836 个任务并混入思维链示例,训练出 Flan-PaLM 540B 与 Flan-T5。论文称 Flan-PaLM 540B 在 MMLU、BBH、TyDiQA、MGSM 等基准上取得 SOTA,Flan-T5-XL(3B)常能追平或超过 PaLM 62B。作者指出,模型越忠实地遵循指令,也就越容易被精心构造的对抗指令利用。
Read originalfailurefirst.org