Model-Driven Labeled Data Free Fine-tuning
Keywords: Fine-tuning, unsupervised learning, foundation models
Abstract: Supervised fine-tuning is a prevalent technique for boosting model performance. However, it heavily depends on extensive training over labeled data. This paper introduces a novel model-driven fine-tuning method that operates independently of supervised training and labeled data. By harnessing the collective intelligence of a diverse model pool, our method enhances individual model performance through a two-phase process. Initially, we consolidate the expertise of the models within the pool to create a general meta-model. This meta-model then serves as a guide for iteratively fine-tuning the original models in a few shots, promoting a synergistic improvement in performance. Our experimental results show that this model-driven approach not only surpasses the performance of full-parameter fine-tuning models but also does so without the need for supervised training. This breakthrough offers a cost-effective and scalable alternative to traditional supervised fine-tuning, addressing the challenge of data scarcity and paving the way for future research in unsupervised model enhancement. Our work represents a significant step towards making fine-tuning techniques more accessible and practical in environments where labeled data is limited or even unavailable.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11370
Loading