The black-box nature of large language models (LLMs) poses challenges in interpreting results, impacting issues such as data intellectual property protection and hallucination tracing. Training data attribution (TDA) methods are considered effective solutions to address these challenges. Most recent TDA methods rely on influence functions, assuming the model achieves minimized empirical risk. However, achieving this criterion is difficult, and sourcing accuracy can be compromised by fitting errors during model training. In this paper, we introduce a novel TDA method called Debias and Denoise Attribution (DDA), which enhances influence functions by addressing fitting errors. Specifically, the debias strategy seeks to improve the performance of influence functions by eliminating the knowledge bias present in the base model before fine-tuning, while the denoise strategy aims to reduce discrepancies in influence scores arising from varying degrees of fitting during the training process through smoothing techniques. Experimental results demonstrate that our method significantly outperforms existing approaches, achieving an averaged AUC of 91.64%. Moreover, DDA exhibits strong generality and scalability across various sources and different-scale models like LLaMA2, QWEN2, and Mistral.

Recent debates raised concerns that language models may favor certain viewpoints. But what if the solution is not to aim for a 'view from nowhere' but rather to leverage different viewpoints? We introduce Plurals, a system and Python library for pluralistic AI deliberation. Plurals consists of Agents (LLMs, optionally with personas) which deliberate within customizable Structures, with Moderators overseeing deliberation. Plurals is a generator of simulated social ensembles. Plurals integrates with government datasets to create nationally representative personas, includes deliberation templates inspired by deliberative democracy, and allows users to customize both information-sharing structures and deliberation behavior within Structures. Six case studies demonstrate fidelity to theoretical constructs and efficacy. Three randomized experiments show simulated focus groups produced output resonant with an online sample of the relevant audiences (chosen over zero-shot generation in 75% of trials). Plurals is both a paradigm and a concrete system for pluralistic AI. The Plurals library is available at https://github.com/josh-ashkinaze/plurals and will be continually updated.

Editor's note: This analysis is part of The Atlantic's investigation into the OpenSubtitles data set. You can access the search tool directly here. Find The Atlantic's search tool for books used to train AI here. For as long as generative-AI chatbots have been on the internet, Hollywood writers have wondered if their work has been used to train them. The chatbots are remarkably fluent with movie references, and companies seem to be training them on all available sources.

LLMs are increasingly fine-tuned using RLHF datasets to align them with human preferences and values. However, very limited research has investigated which specific human values are operationalized through these datasets. In this paper, we introduce Value Imprint, a framework for auditing and classifying the human values embedded within RLHF datasets. To investigate the viability of this framework, we conducted three case study experiments by auditing the Anthropic/hh-rlhf, OpenAI WebGPT Comparisons, and Alpaca GPT-4-LLM datasets to examine the human values embedded within them. Our analysis involved a two-phase process. During the first phase, we developed a taxonomy of human values through an integrated review of prior works from philosophy, axiology, and ethics. Then, we applied this taxonomy to annotate 6,501 RLHF preferences. During the second phase, we employed the labels generated from the annotation as ground truth data for training a transformer-based machine learning model to audit and classify the three RLHF datasets. Through this approach, we discovered that information-utility values, including Wisdom/Knowledge and Information Seeking, were the most dominant human values within all three RLHF datasets. In contrast, prosocial and democratic values, including Well-being, Justice, and Human/Animal Rights, were the least represented human values. These findings have significant implications for developing language models that align with societal values and norms. We contribute our datasets to support further research in this area.

With fairness concerns gaining significant attention in Machine Learning (ML), several bias mitigation techniques have been proposed, often compared against each other to find the best method. These benchmarking efforts tend to use a common setup for evaluation under the assumption that providing a uniform environment ensures a fair comparison. However, bias mitigation techniques are sensitive to hyperparameter choices, random seeds, feature selection, etc., meaning that comparison on just one setting can unfairly favour certain algorithms. In this work, we show significant variance in fairness achieved by several algorithms and the influence of the learning pipeline on fairness scores. We highlight that most bias mitigation techniques can achieve comparable performance, given the freedom to perform hyperparameter optimization, suggesting that the choice of the evaluation parameters-rather than the mitigation technique itself-can sometimes create the perceived superiority of one method over another. We hope our work encourages future research on how various choices in the lifecycle of developing an algorithm impact fairness, and trends that guide the selection of appropriate algorithms.

Open-source large language models are becoming increasingly available and popular among researchers and practitioners. While significant progress has been made on open-weight models, open training data is a practice yet to be adopted by the leading open-weight models creators. At the same time, there researchers are working to make language models safer. We propose a data curation pipeline to reduce harmful outputs by models trained on public domain data. There are unique challenges to working with public domain data, as these sources differ from web text in both form and content. Many sources are historical documents and are the result of Optical Character Recognition (OCR). Consequently, current state-of-the-art approaches to toxicity filtering are often infeasible or inappropriate for open data models. In this paper, we introduce a new fully open-source pipeline for open-data toxicity filtering. Our contributions are threefold. We create a custom training dataset, ToxicCommons, which is composed of texts which have been classified across five different dimensions (racial/origin-based, gender/sex-based, religious, ability-based discrimination, and violence). We use this dataset to train a custom classifier, Celadon, that can be used to detect toxic content in open data more efficiently at a larger scale. Finally, we describe the balanced approach to content filtration that optimizes safety filtering with respect to the filtered data available for training.

For all the ways that AI might have transformed the world, I doubt many people expected flirty serial killer chatbots to be part of it. Yet on the seedier corners of the internet, there are scores of AIs built specifically to give some most vicious killers in history a romantic twist. To see just how dark these bots could be, I decided to step up for a spot of serial killer speed dating. I went on six'dates' with some of history's most notorious murderers and, perhaps unsurprisingly, it didn't go great. Character.ai is the subject of a lawsuit alleging that its bots drove a 14-year-old boy to take his life and has been used to host ghoulish replicas of murder victims such as Brianna Ghey. Across my six dates on the site, I found myself threatened with violence, stalked, invited to remote locations, and generally met with some extraordinarily uncomfortable flirting.

Federated learning facilitates collaborative machine learning, enabling multiple participants to collectively develop a shared model while preserving the privacy of individual data. The growing importance of the "right to be forgotten" calls for effective mechanisms to facilitate data removal upon request. In response, federated unlearning (FU) has been developed to efficiently eliminate the influence of specific data from the model. Current FU methods primarily rely on approximate unlearning strategies, which seek to balance data removal efficacy with computational and communication costs, but often fail to completely erase data influence. To address these limitations, we propose FedUHB, a novel exact unlearning approach that leverages the Polyak heavy ball optimization technique, a first-order method, to achieve rapid retraining. In addition, we introduce a dynamic stopping mechanism to optimize the termination of the unlearning process. Our extensive experiments show that FedUHB not only enhances unlearning efficiency but also preserves robust model performance after unlearning. Furthermore, the dynamic stopping mechanism effectively reduces the number of unlearning iterations, conserving both computational and communication resources. FedUHB can be proved as an effective and efficient solution for exact data removal in federated learning settings.

Federated unlearning (FU) offers a promising solution to effectively address the need to erase the impact of specific clients' data on the global model in federated learning (FL), thereby granting individuals the ``Right to be Forgotten". The most straightforward approach to achieve unlearning is to train the model from scratch, excluding clients who request data removal, but it is resource-intensive. Current state-of-the-art FU methods extend traditional FL frameworks by leveraging stored historical updates, enabling more efficient unlearning than training from scratch. However, the use of stored updates introduces significant privacy risks. Adversaries with access to these updates can potentially reconstruct clients' local data, a well-known vulnerability in the privacy domain. While privacy-enhanced techniques exist, their applications to FU scenarios that balance unlearning efficiency with privacy protection remain underexplored. To address this gap, we propose FedADP, a method designed to achieve both efficiency and privacy preservation in FU. Our approach incorporates an adaptive differential privacy (DP) mechanism, carefully balancing privacy and unlearning performance through a novel budget allocation strategy tailored for FU. FedADP also employs a dual-layered selection process, focusing on global models with significant changes and client updates closely aligned with the global model, reducing storage and communication costs. Additionally, a novel calibration method is introduced to facilitate effective unlearning. Extensive experimental results demonstrate that FedADP effectively manages the trade-off between unlearning efficiency and privacy protection.

Sparse Autoencoders for transformer-based language models are typically defined independently per layer. In this work we analyze statistical relationships between features in adjacent layers to understand how features evolve through a forward pass. We provide a graph visualization interface for features and their most similar next-layer neighbors, and build communities of related features across layers. We find that a considerable amount of features are passed through from a previous layer, some features can be expressed as quasi-boolean combinations of previous features, and some features become more specialized in later layers.