In 2019, the paper entitled "Model Cards for Model Reporting" introduced a new tool for documenting model performance and encouraged the practice of transparent reporting for a defined list of categories. One of the categories detailed in that paper is ethical considerations, which includes the subcategories of data, human life, mitigations, risks and harms, and use cases. We propose to reclassify this category in the original model card due to the recent maturing of the field known as trustworthy AI, a term which analyzes whether the algorithmic properties of the model indicate that the AI system is deserving of trust from its stakeholders. In our examination of trustworthy AI, we highlight three respected organizations - the European Commission's High-Level Expert Group on AI, the OECD, and the U.S.-based NIST - that have written guidelines on various aspects of trustworthy AI. These recent publications converge on numerous characteristics of the term, including accountability, explainability, fairness, privacy, reliability, robustness, safety, security, and transparency, while recognizing that the implementation of trustworthy AI varies by context. Our reclassification of the original model-card category known as ethical considerations involves a two-step process: 1) adding a new category known as trustworthiness, where the subcategories will be derived from the discussion of trustworthy AI in our paper, and 2) maintaining the subcategories of ethical considerations under a renamed category known as risk environment and risk management, a title which we believe better captures today's understanding of the essence of these topics. We hope that this reclassification will further the goals of the original paper and continue to prompt those releasing trained models to accompany these models with documentation that will assist in the evaluation of their algorithmic properties.

We explore machine unlearning (MU) in the domain of large language models (LLMs), referred to as LLM unlearning. This initiative aims to eliminate undesirable data influence (e.g., sensitive or illegal information) and the associated model capabilities, while maintaining the integrity of essential knowledge generation and not affecting causally unrelated information. We envision LLM unlearning becoming a pivotal element in the life-cycle management of LLMs, potentially standing as an essential foundation for developing generative AI that is not only safe, secure, and trustworthy, but also resource-efficient without the need of full retraining. We navigate the unlearning landscape in LLMs from conceptual formulation, methodologies, metrics, and applications. In particular, we highlight the often-overlooked aspects of existing LLM unlearning research, e.g., unlearning scope, data-model interaction, and multifaceted efficacy assessment. We also draw connections between LLM unlearning and related areas such as model editing, influence functions, model explanation, adversarial training, and reinforcement learning. Furthermore, we outline an effective assessment framework for LLM unlearning and explore its applications in copyright and privacy safeguards and sociotechnical harm reduction.

We explored cultural biases-individualism vs. collectivism-in ChatGPT across three Western languages (i.e., English, German, and French) and three Eastern languages (i.e., Chinese, Japanese, and Korean). When ChatGPT adopted an individualistic persona in Western languages, its collectivism scores (i.e., out-group values) exhibited a more negative trend, surpassing their positive orientation towards individualism (i.e., in-group values). Conversely, when a collectivistic persona was assigned to ChatGPT in Eastern languages, a similar pattern emerged with more negative responses toward individualism (i.e., out-group values) as compared to collectivism (i.e., in-group values). The results indicate that when imbued with a particular social identity, ChatGPT discerns in-group and out-group, embracing in-group values while eschewing out-group values. Notably, the negativity towards the out-group, from which prejudices and discrimination arise, exceeded the positivity towards the in-group. The experiment was replicated in the political domain, and the results remained consistent. Furthermore, this replication unveiled an intrinsic Democratic bias in Large Language Models (LLMs), aligning with earlier findings and providing integral insights into mitigating such bias through prompt engineering. Extensive robustness checks were performed using varying hyperparameter and persona setup methods, with or without social identity labels, across other popular language models.

Living things, computers, societies, and even books are part of a grand evolutionary struggle to survive. That struggle shapes nature, nations, religions, art, science, and you. What you think, feel, and do is determined by it. Darwinian evolution does not apply solely to the genes that are stored in DNA. Using the insights of Alan Turing and Richard Dawkins, we will see that it also applies to the memes we store in our brains and the information we store in our computers. The next time you run for president, fight a war, or just deal with the ordinary problems humans are heir to, perhaps this book will be of use. If you want to understand why and when you will die, or if you want to achieve greatness this book may help. If you are concerned about where the computer revolution is headed, this book may provide some answers.

The rise of large language models (LLMs) has drawn attention to the existence of "jailbreaks" that allow the models to be used maliciously. However, there is no standard benchmark for measuring the severity of a jailbreak, leaving authors of jailbreak papers to create their own. We show that these benchmarks often include vague or unanswerable questions and use grading criteria that are biased towards overestimating the misuse potential of low-quality model responses. Some jailbreak techniques make the problem worse by decreasing the quality of model responses even on benign questions: we show that several jailbreaking techniques substantially reduce the zero-shot performance of GPT-4 on MMLU. Jailbreaks can also make it harder to elicit harmful responses from an "uncensored" open-source model. We present a new benchmark, StrongREJECT, which better discriminates between effective and ineffective jailbreaks by using a higher-quality question set and a more accurate response grading algorithm. We show that our new grading scheme better accords with human judgment of response quality and overall jailbreak effectiveness, especially on the sort of low-quality responses that contribute the most to over-estimation of jailbreak performance on existing benchmarks. We release our code and data at https://github.com/alexandrasouly/strongreject.

We study the per-datum Membership Inference Attacks (MIAs), where an attacker aims to infer whether a fixed target datum has been included in the input dataset of an algorithm and thus, violates privacy. First, we define the membership leakage of a datum as the advantage of the optimal adversary targeting to identify it. Then, we quantify the per-datum membership leakage for the empirical mean, and show that it depends on the Mahalanobis distance between the target datum and the data-generating distribution. We further assess the effect of two privacy defences, i.e. adding Gaussian noise and sub-sampling. We quantify exactly how both of them decrease the per-datum membership leakage. Our analysis builds on a novel proof technique that combines an Edgeworth expansion of the likelihood ratio test and a Lindeberg-Feller central limit theorem. Our analysis connects the existing likelihood ratio and scalar product attacks, and also justifies different canary selection strategies used in the privacy auditing literature. Finally, our experiments demonstrate the impacts of the leakage score, the sub-sampling ratio and the noise scale on the per-datum membership leakage as indicated by the theory.

While displaying impressive generation capabilities across many tasks, Large Language Models (LLMs) still struggle with crucial issues of privacy violation and unwanted exposure of sensitive data. This raises an essential question: how should we prevent such undesired behavior of LLMs while maintaining their strong generation and natural language understanding (NLU) capabilities? In this work, we introduce a novel approach termed deliberate imagination in the context of LLM unlearning. Instead of trying to forget memorized data, we employ a self-distillation framework, guiding LLMs to deliberately imagine alternative scenarios. As demonstrated in a wide range of experiments, the proposed method not only effectively unlearns targeted text but also preserves the LLMs' capabilities in open-ended generation tasks as well as in NLU tasks. Our results demonstrate the usefulness of this approach across different models and sizes, and also with parameter-efficient fine-tuning, offering a novel pathway to addressing the challenges with private and sensitive data in LLM applications.

Large Language Models have proven highly successful at modelling a variety of tasks. However, this comes at a steep computational cost that hinders wider industrial uptake. In this pa005 per, we present MWT: a Multi-Word Tokenizer that goes beyond word boundaries by representing frequent multi-word expressions as single tokens. MWTs produce a more compact and efficient tokenization that yields two benefits: (1) Increase in performance due to a greater coverage of input data given a fixed sequence length and budget; (2) Faster and lighter inference due to the ability to reduce the sequence length with negligible drops in performance. Our results show that MWT is more robust across shorter sequence lengths, thus allowing for major speedups via early sequence truncation.

Albeit Natural Language Processing has seen major breakthroughs in the last few years, transferring such advances into real-world business cases can be challenging. One of the reasons resides in the displacement between popular benchmarks and actual data. Lack of supervision, unbalanced classes, noisy data and long documents often affect real problems in vertical domains such as finance, law and health. To support industry-oriented research, we present BUSTER, a BUSiness Transaction Entity Recognition dataset. The dataset consists of 3779 manually annotated documents on financial transactions. We establish several baselines exploiting both general-purpose and domain-specific language models. The best performing model is also used to automatically annotate 6196 documents, which we release as an additional silver corpus to BUSTER.

How can we detect if copyrighted content was used in the training process of a language model, considering that the training data is typically undisclosed? We are motivated by the premise that a language model is likely to identify verbatim excerpts from its training text. We propose DE-COP, a method to determine whether a piece of copyrighted content was included in training. DE-COP's core approach is to probe an LLM with multiple-choice questions, whose options include both verbatim text and their paraphrases. We construct BookTection, a benchmark with excerpts from 165 books published prior and subsequent to a model's training cutoff, along with their paraphrases. Our experiments show that DE-COP surpasses the prior best method by 9.6% in detection performance (AUC) on models with logits available. Moreover, DE-COP also achieves an average accuracy of 72% for detecting suspect books on fully black-box models where prior methods give $\approx$ 4% accuracy. Our code and datasets are available at https://github.com/avduarte333/DE-COP_Method