Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

Detecting Distillation Data from Reasoning Models

Created by
  • Haebom

Author

Hengxiang Zhang, Hyeong Kyu Choi, Sharon Li, Hongxin Wei

Outline

This paper presents research to address the problem of evaluation data contamination that can arise during the distillation process, which improves the inference capability of large-scale language models. Specifically, considering the partial availability of distillation data, we define the task of distillation data detection and propose a Token Probability Deviation (TBD) method that utilizes token generation probability patterns. TBD analyzes the tendency of a distilled model to generate certain tokens with high probability for previously learned questions and with low probability for new questions. This method detects distillation data by measuring the deviation in token probability. Experimental results show that TBD achieves an AUC of 0.918 and a TPR of 0.470 at a FPR of 1% on the S1 dataset.

Takeaways, Limitations

•
Takeaways:
◦
A new methodology is presented to address the problem of evaluation data contamination during the distillation process.
◦
Proposing an effective data distillation detection methodology utilizing token generation probability patterns.
◦
Setting a realistic problem considering the availability of partial distillation data
◦
High-performance distillation data detection results (AUC 0.918, TPR@1% FPR 0.470)
•
Limitations:
◦
Performance evaluation is only performed on a specific dataset (S1).
◦
Further validation of generalizability to other types of models and datasets is needed.
◦
Further analysis of the computational complexity and efficiency of the TBD method is needed.
👍
Made with Slashpage