Purifying Large Language Models by Ensembling a Small Language Model (2402.14845v1)

Abstract: The emerging success of LLMs heavily relies on collecting abundant training data from external (untrusted) sources. Despite substantial efforts devoted to data cleaning and curation, well-constructed LLMs have been reported to suffer from copyright infringement, data poisoning, and/or privacy violations, which would impede practical deployment of LLMs. In this study, we propose a simple and easily implementable method for purifying LLMs from the negative effects caused by uncurated data, namely, through ensembling LLMs with benign and small LLMs (SLMs). Aside from theoretical guarantees, we perform comprehensive experiments to empirically confirm the efficacy of ensembling LLMs with SLMs, which can effectively preserve the performance of LLMs while mitigating issues such as copyright infringement, data poisoning, and privacy violations.

• The paper introduces an ensemble method that fuses LLMs and SLMs to purify models from uncurated data issues without altering their parameters.
• It leverages output probability blending via the CP-ΔΔ algorithm to significantly reduce copyright infringements, data poisoning, and privacy leaks.
• The study demonstrates a minimal performance trade-off and flexible integration with other enhancement techniques for real-world LLM applications.

Ensembling Large and Small LLMs for Mitigating Uncurated Data Effects

Introduction

The development and deployment of LLMs have undergone substantial progress, predominantly fueled by extensive web-collected training datasets. However, these datasets often include uncurated content leading to legal, ethical, and privacy concerns such as copyright infringement, data poisoning, and personal identifiable information (PII) leakage. Traditional data curation methods are labor-intensive and thus, not entirely effective in mitigating these issues. In this context, the proposed paper introduces a simpler, more feasible alternative—a model ensemble strategy that combines the capabilities of an untrusted LLM with a benign Small LLM (SLM) to purify the LLM from the adverse impacts of uncurated data.

The Ensemble Strategy

The core proposition of this paper is the ensemble method, which melds the output probabilities of both an LLM and an SLM. This blend aims to leverage the high performance of LLMs while utilizing the benign, well-curated nature of SLMs to counteract the negative effects of untrusted data in LLMs. The ensemble operates at the logits level, implying a plug-and-play approach that does not necessitate alterations to the original models' parameters. It leverages the CP-ΔΔ algorithm under the premise that both models fulfill a "sharded-safe" function, which is crucial for implementing copyright protection measures. The method's flexibility is evident in its capacity to adjust ensemble weights dynamically, offering a spectrum of models with varying degrees of purification and performance.

Evaluation and Findings

The evaluation of this ensemble strategy across nine different LLMs on ten benchmark datasets demonstrates its efficacy in significantly reducing copyright infringements, data poisonings, and privacy violations. Notably, the method shows a minimal trade-off between purification of the model and retention of standard performance. This indicates the ensemble's potential as an efficient and flexible solution to align with changing standards and regulations, especially in the copyright domain. Furthermore, experiments highlight that the ensemble method can be seamlessly integrated with other model enhancement strategies, showcasing its broad applicability.

Practical Implications and Future Directions

From a practical standpoint, this ensemble approach presents a promising avenue for mitigating the risks associated with uncurated data in LLMs without compromising their performance. Its ability to dynamically adjust ensemble weights opens doors to real-time model optimization in response to varying operational requirements and regulations. Moreover, the paper hints at the potential for further exploration into the synergistic integration of this ensemble strategy with other model enhancement and acceleration techniques, potentially broadening its utility beyond data purification alone.

Conclusion

This research delineates a novel and effective method for purifying LLMs from the repercussions of uncurated data through a strategic ensemble with SLMs. The strategy not only preserves the LLMs' performance but also mitigates legal and ethical risks, representing a significant stride towards the responsible use of LLMs. Given the ever-increasing reliance on LLMs across various domains, such a method holds considerable promise for enhancing the utility and acceptability of LLM applications in real-world scenarios, setting a precedent for future research in this critical area of AI ethics and governance.