Skip to content

Settings for instruction-tuning #43

@KaiLv69

Description

@KaiLv69

Hi, Adan是一个性能十分优秀的优化器,谢谢你们的工作。

但我最近在尝试用Adan进行指令微调时,发现loss曲线很漂亮,但是下游任务表现(GSM-8k)不如预期。
同样的数据处理和评测,AdamW大概9.63,Adan只有5.08左右。

AdamW超参数:weight_decay 0.01, lr 2e-5
Adan超参数:weight_decay 0.02,按照repo的建议lr尝试了2e-4 1e-4, GSM8k都比较低
lr scheduler都是3%升到最高然后下降到0

AdamW的训练loss曲线:
image

Adan的训练loss曲线:
image

使用的代码:

from adan import Adan
optimizer = Adan(model.parameters(), lr=args.lr, weight_decay=0.02, foreach=True, fused=True)

想知道有没有一些对指令微调的超参设置建议?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions