知识卡片
协作融合须检验任务间干扰
多个团队从同一基座微调再汇总,有机会复用分散任务经验,也可能因目标冲突相互抵消。应保存基座版本、各增量与评估集,比较合成后的各任务表现,而非只给新基座一个综合分。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 进阶微调技术
原书出处 · 1 处
- · 查看原书依据
all across the organization, who take the model and fine-tune it on their own tasks. They then upload the fine-tuned models back. You can then merge the weights of all the models, resulting in a stronger pre-trained model. This model can then be used as a new version of the base model. This process has been coined Collaborative Descent (ColD) Fusion by Don-Yehiya et al.
参考来源
- 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html)
- 结论依据:摘录中关于协作融合须检验任务间干扰的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:all across the organization, who take the model and fine-tune it on their own tasks. They then upload the fine-tuned models back. You can then merge the weights of all the models, resulting in a stronger pre-trained model. This model can then be used as a new version of the base model. This process has been coined Collaborative Descent (ColD) Fusion by Don-Yehiya et al.