LoRA-BAM: Input Filtering for Fine-tuned LLMs via Boxed Abstraction Monitors over LoRA Layers
This addresses overfitting issues in domain-specific fine-tuning for users of LLMs, though it is incremental as it builds on existing LoRA and OoD detection techniques.
The paper tackles the problem of fine-tuned large language models becoming unreliable on out-of-distribution queries by proposing LoRA-BAM, a method that uses boxed abstraction monitors over LoRA layers to filter such queries, resulting in lightweight and interpretable OoD detection.
Fine-tuning large language models (LLMs) improves performance on domain-specific tasks but can lead to overfitting, making them unreliable on out-of-distribution (OoD) queries. We propose LoRA-BAM - a method that adds OoD detection monitors to the LoRA layer using boxed abstraction to filter questions beyond the model's competence. Feature vectors from the fine-tuning data are extracted via the LLM and clustered. Clusters are enclosed in boxes; a question is flagged as OoD if its feature vector falls outside all boxes. To improve interpretability and robustness, we introduce a regularization loss during fine-tuning that encourages paraphrased questions to stay close in the feature space, and the enlargement of the decision boundary is based on the feature variance within a cluster. Our method complements existing defenses by providing lightweight and interpretable OoD detection.