CDD Technique for Recovering Original LLM Fine-Tuning Data Unveiled

Researchers have proposed Contrastive Decoding Diffing (CDD), a method that recovers the exact original text of data used in fine-tuning by leveraging the Logit difference between a fine-tuned model and its base model, even in a 'Grey-box' environment where there is no access to the model's weights or activations.

Unlike the existing Activation Difference Lens (ADL) approach, which requires full access to model weights and can only recover vague, domain-level information, CDD has the following characteristics.

• High recovery accuracy: In SDF benchmark tests, across various model families ranging from 1B to 32B parameters, CDD achieved a recovery score of 4/5 or higher in 19 out of 20 pairs.
• Efficiency: It operates with a single default configuration, without requiring separate model-specific calibration or layer-selection processes.
• Data contamination detection: During the research, it was discovered that a specific person's name ('Dr. Elena Rodriguez') was commonly recovered even across fine-tuning data from different domains. This suggests that models such as Claude 3.5 Sonnet tend to overuse certain personas when generating synthetic data, and that this can be reflected in fine-tuning data and become imprinted in the model.