Machine Unlearning for Large Language Models, referred to as LLM Unlearning is getting more and more attention as a result of regurgitation of sensitive and harmful content. In this paper, we present our method architecture, results, and analysis of our submission to Task4: Unlearning sensitive content from Large Language Models. This task includes three subtasks of LLM Unlearning on 1) Long Synthetic documents, 2) Short Synthetic documents, and 3) Real Training documents. Getting rid of the impact of undesirable and unauthorized responses is the core objective of unlearning. Furthermore, it is expected that unlearning should not have an adverse impact on the usability of the model. In this paper, we provide an approach for LLM unlearning that tries to make the model forget while maintaining usability of the model. We perform adaptive weight tuning with Gradient Ascent, KL minimization and Gradual Negative Matching loss functions. Our submission balances retain and forget abilities of the model while outperforming provided benchmarks.

Social media are heavily used by many users to share their mental health concerns and diagnoses. This trend has turned social media into a large-scale resource for researchers focused on detecting mental health conditions. Social media usage varies considerably across individuals. Thus, classification of patterns, including detecting signs of depression, must account for such variation. We address the disparity in classification effectiveness for users with little activity (e.g., new users). Our evaluation, performed on a large-scale dataset, shows considerable detection discrepancy based on user posting frequency. For instance, the F1 detection score of users with an above-median versus below-median number of posts is greater than double (0.803 vs 0.365) using a conventional CNN-based model; similar results were observed on lexical and transformer-based classifiers. To complement this evaluation, we propose a dynamic thresholding technique that adjusts the classifier’s sensitivity as a function of the number of posts a user has. This technique alone reduces the margin between users with many and few posts, on average, by 45% across all methods and increases overall performance, on average, by 33%. These findings emphasize the importance of evaluating and tuning natural language systems for potentially vulnerable populations.