Selective Fine-Tuning for Targeted and Robust Concept Unlearning

Text guided diffusion models are used by millions of users, but can be easily exploited to produce harmful content. Concept unlearning methods aim at reducing the models' likelihood of generating harmful content. Traditionally, this has been tackled at an individual concept level, with only a handful of recent works considering more realistic concept combinations. However, state of the art methods depend on full finetuning, which is computationally expensive. Concept localisation methods can facilitate selective finetuning, but existing techniques are static, resulting in suboptimal utility. In order to tackle these challenges, we propose TRUST (Targeted Robust Selective fine Tuning), a novel approach for dynamically estimating target concept neurons and unlearning them through selective finetuning, empowered by a Hessian based regularization. We show experimentally, against a number of SOTA baselines, that TRUST is robust against adversarial prompts, preserves generation quality to a significant degree, and is also significantly faster than the SOTA. Our method achieves unlearning of not only individual concepts but also combinations of concepts and conditional concepts, without any specific regularization.

AmalREC: A Dataset for Relation Extraction and Classification Leveraging Amalgamation of Large Language Models

Existing datasets for relation classification and extraction often exhibit limitations such as restricted relation types and domain-specific biases. This work presents a generic framework to generate well-structured sentences from given tuples with the help of Large Language Models (LLMs). This study has focused on the following major questions: (i) how to generate sentences from relation tuples, (ii) how to compare and rank them, (iii) can we combine strengths of individual methods and amalgamate them to generate an even bette quality of sentences, and (iv) how to evaluate the final dataset? For the first question, we employ a multifaceted 5-stage pipeline approach, leveraging LLMs in conjunction with template-guided generation. We introduce Sentence Evaluation Index(SEI) that prioritizes factors like grammatical correctness, fluency, human-aligned sentiment, accuracy, and complexity to answer the first part of the second question. To answer the second part of the second question, this work introduces a SEI-Ranker module that leverages SEI to select top candidate generations. The top sentences are then strategically amalgamated to produce the final, high-quality sentence. Finally, we evaluate our dataset on LLM-based and SOTA baselines for relation classification. The proposed dataset features 255 relation types, with 15K sentences in the test set and around 150k in the train set organized in, significantly enhancing relational diversity and complexity. This work not only presents a new comprehensive benchmark dataset for RE/RC task, but also compare different LLMs for generation of quality sentences from relational tuples.