Extracting alignment data in open models

TL;DR

This paper demonstrates that significant alignment training data can be extracted from post-trained models using embedding-based similarity measures, revealing potential risks and implications for model training and distillation practices.

Contribution

It introduces a novel method for extracting alignment data via embeddings, showing that models regurgitate training data from post-training phases and that this data can be used to recover original performance.

Findings

Embedding models outperform string matching in data extraction.

Models readily regurgitate training data from post-training phases.

Extracted data can be used to retrain models, recovering performance.

Abstract

In this work, we show that it is possible to extract significant amounts of alignment training data from a post-trained model -- useful to steer the model to improve certain capabilities such as long-context reasoning, safety, instruction following, and maths. While the majority of related work on memorisation has focused on measuring success of training data extraction through string matching, we argue that embedding models are better suited for our specific goals. Distances measured through a high quality embedding model can identify semantic similarities between strings that a different metric such as edit distance will struggle to capture. In fact, in our investigation, approximate string matching would have severely undercounted (by a conservative estimate of $10\times$) the amount of data that can be extracted due to trivial artifacts that deflate the metric. Interestingly, we find that models readily regurgitate training data that was used in post-training phases such as SFT or RL. We show that this data can be then used to train a base model, recovering a meaningful amount of the original performance. We believe our work exposes a possibly overlooked risk towards extracting alignment data. Finally, our work opens up an interesting discussion on the downstream effects of distillation practices: since models seem to be regurgitating aspects of their training set, distillation can therefore be thought of as indirectly training on the model's original dataset.