Skip to content

C3 finding: V-PEFT shows dataset-dependent accuracy and compute-efficiency trade-offs #277

Description

@caffeine-2026

Context

在 C3 工业缺陷小样本实践中,我们在 NEU-DET 与 DeepPCB 上完成了 Full-SFT、Frozen Backbone 与 V-PEFT 的同预算、多 seed 对照,并扩展到 10/50/100/500 张 nested few-shot scaling。

Observed Results

  • V-PEFT reduces trainable parameters by 76.32%.
  • GPU memory saving is only about 1.14%-1.53%.
  • Training time is 5.67%-19.57% higher than Full-SFT under the tested setup.
  • NEU-DET preserves Full-SFT accuracy substantially better than DeepPCB.
  • DeepPCB retention improves as sample size grows.

这些是当前实验条件下的技术观察,不将其描述为 YOLO-Master bug,也不主张 V-PEFT 是 universal winner。

Question

希望进一步讨论:

  1. LoRA target placement 是否影响显存/时间效率;
  2. rank / target / planner budget 是否需要数据集自适应;
  3. 是否可以减少 activation / optimizer / adapter overhead;
  4. 是否值得增加 adaptive V-PEFT policy。

Evidence

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions