Search for: All records

Instruction tuning is critical for adapting large language models (LLMs) to downstream tasks, and recent studies have demonstrated that small amounts of human-curated data can outperform larger datasets, challenging traditional data scaling laws. While LLM-based data quality rating systems offer a cost-effective alternative to human annotation, they often suffer from inaccuracies and biases, even in powerful models like GPT-4. In this work, we introduce DS2, a Diversity-aware Score curation method for Data Selection. By systematically modeling error patterns through a score transition matrix, DS2 corrects LLM-based scores and promotes diversity in the selected data samples. Our approach shows that a curated subset (just 3.3% of the original dataset) outperforms full-scale datasets (300k samples) across various machine-alignment benchmarks, and matches or surpasses human-aligned datasets such as LIMA with the same sample size (1k samples). These findings challenge conventional data scaling assumptions, highlighting that redundant, low-quality samples can degrade performance and reaffirming that "more can be less." 

Large language models (LLMs) have advanced to encompass extensive knowledge across diverse domains. Yet controlling what a large language model should not know is important for ensuring alignment and thus safe use. However, accurately and efficiently unlearning knowledge from an LLM remains challenging due to the potential collateral damage caused by the fuzzy boundary between retention and forgetting, and the large computational requirements for optimization across state-of-the-art models with hundreds of billions of parameters. In this work, we present \textbf{Embedding-COrrupted (ECO) Prompts}, a lightweight unlearning framework for large language models to address both the challenges of knowledge entanglement and unlearning efficiency. Instead of relying on the LLM itself to unlearn, we enforce an unlearned state during inference by employing a prompt classifier to identify and safeguard prompts to forget. We learn corruptions added to prompt embeddings via zeroth order optimization toward the unlearning objective offline and corrupt prompts flagged by the classifier during inference. We find that these embedding-corrupted prompts not only lead to desirable outputs that satisfy the unlearning objective but also closely approximate the output from a model that has never been trained on the data intended for forgetting. Through extensive experiments on unlearning, we demonstrate the superiority of our method in achieving promising unlearning at \textit{nearly zero side effects} in general domains and domains closely related to the unlearned ones. Additionally, we highlight the scalability of our method to 100 LLMs, ranging from 0.5B to 236B parameters, incurring no additional cost as the number of parameters increases. We have made our code publicly available at \url{this https URL}. 

Federated learning (FL) is a learning paradigm that allows the central server to learn from different data sources while keeping the data private locally. Without controlling and monitoring the local data collection process, the locally available training labels are likely noisy, i.e., the collected training labels differ from the unobservable ground truth. Additionally, in heterogenous FL, each local client may only have access to a subset of label space (referred to as openset label learning), meanwhile without overlapping with others. In this work, we study the challenge of FL with local openset noisy labels. We observe that many existing solutions in the noisy label literature, e.g., loss correction, are ineffective during local training due to overfitting to noisy labels and being not generalizable to openset labels. For the methods in FL, different estimated metrics are shared. To address the problems, we design a label communication mechanism that shares "contrastive labels" randomly selected from clients with the server. The privacy of the shared contrastive labels is protected by label differential privacy (DP). Both the DP guarantee and the effectiveness of our approach are theoretically guaranteed. Compared with several baseline methods, our solution shows its efficiency in several public benchmarks and real-world datasets under different noise ratios and noise models. 

Performative prediction, as introduced by Perdomo et al, is a framework for studying social prediction in which the data distribution itself changes in response to the deployment of a model. Existing work in this field usually hinges on three assumptions that are easily violated in practice: that the performative risk is convex over the deployed model, that the mapping from the model to the data distribution is known to the model designer in advance, and the first-order information of the performative risk is available. In this paper, we initiate the study of performative prediction problems that do not require these assumptions. Specifically, we develop a reparameterization framework that reparametrizes the performative prediction objective as a function of the induced data distribution. We then develop a two-level zeroth-order optimization procedure, where the first level performs iterative optimization on the distribution parameter space, and the second level learns the model that induces a particular target distribution at each iteration. Under mild conditions, this reparameterization allows us to transform the non-convex objective into a convex one and achieve provable regret guarantees. In particular, we provide a regret bound that is sublinear in the total number of performative samples taken and is only polynomial in the dimension of the model parameter. 

Language models have shown promise in various tasks but can be affected by undesired data during training, fine-tuning, or alignment. For example, if some unsafe conversations are wrongly annotated as safe ones, the model fine-tuned on these samples may be harmful. Therefore, the correctness of annotations, i.e., the credibility of the dataset, is important. This study focuses on the credibility of real-world datasets, including the popular benchmarks Jigsaw Civil Comments, Anthropic Harmless & Red Team, PKU BeaverTails & SafeRLHF, that can be used for training a harmless language model. Given the cost and difficulty of cleaning these datasets by humans, we introduce a systematic framework for evaluating the credibility of datasets, identifying label errors, and evaluating the influence of noisy labels in the curated language data, specifically focusing on unsafe comments and conversation classification. With the framework, we find and fix an average of 6.16% label errors in 11 datasets constructed from the above benchmarks. The data credibility and downstream learning performance can be remarkably improved by directly fixing label errors, indicating the significance of cleaning existing real-world datasets. 

We reveal and address the frequently overlooked yet important issue of disguised procedural unfairness, namely, the potentially inadvertent alterations on the behavior of neutral (i.e., not problematic) aspects of data generating process, and/or the lack of procedural assurance of the greatest benefit of the least advantaged individuals. Inspired by John Rawls's advocacy for pure procedural justice, we view automated decision-making as a microcosm of social institutions, and consider how the data generating process itself can satisfy the requirements of procedural fairness. We propose a framework that decouples the objectionable data generating components from the neutral ones by utilizing reference points and the associated value instantiation rule. Our findings highlight the necessity of preventing disguised procedural unfairness, drawing attention not only to the objectionable data generating components that we aim to mitigate, but also more importantly, to the neutral components that we intend to keep unaffected.