Abstract:The data mixture for large language model pre-training significantly impacts performance, yet how to determine an effective mixture remains unclear. We propose RegMix to automatically identify a high-performing data mixture by formulating it as a regression task. RegMix involves training a set of small models with diverse data mixtures and fitting a regression model to predict their performance given their respective mixtures. With the fitted regression model, we simulate the top-ranked mixture and use it to train a large-scale model with orders of magnitude more compute. To empirically validate RegMix, we train 512 models with 1M parameters for 1B tokens of different mixtures to fit the regression model and find the optimal mixture. Using this mixture we train a 1B parameter model for 25B tokens (i.e. 1000x larger and 25x longer) which we find performs best among 64 candidate 1B parameter models with other mixtures. Further, our method demonstrates superior performance compared to human selection and achieves results that match or surpass DoReMi, while utilizing only 10% of the compute budget. Our experiments also show that (1) Data mixtures significantly impact performance with single-task performance variations of up to 14.6%; (2) Web corpora rather than data perceived as high-quality like Wikipedia have the strongest positive correlation with downstream performance; (3) Domains interact in complex ways often contradicting common sense, thus automatic approaches like RegMix are needed; (4) Data mixture effects transcend scaling laws, and our approach captures the complexity by considering all domains together. Our code is available at https://github.com/sail-sg/regmix.
Abstract:Tackling non-IID data is an open challenge in federated learning research. Existing FL methods, including robust FL and personalized FL, are designed to improve model performance without consideration of interpreting non-IID across clients. This paper aims to design a novel FL method to robust and interpret the non-IID data across clients. Specifically, we interpret each client's dataset as a mixture of conceptual vectors that each one represents an interpretable concept to end-users. These conceptual vectors could be pre-defined or refined in a human-in-the-loop process or be learnt via the optimization procedure of the federated learning system. In addition to the interpretability, the clarity of client-specific personalization could also be applied to enhance the robustness of the training process on FL system. The effectiveness of the proposed method have been validated on benchmark datasets.
Abstract:Recently, Anil et al. (2024) show that many-shot (up to hundreds of) demonstrations can jailbreak state-of-the-art LLMs by exploiting their long-context capability. Nevertheless, is it possible to use few-shot demonstrations to efficiently jailbreak LLMs within limited context sizes? While the vanilla few-shot jailbreaking may be inefficient, we propose improved techniques such as injecting special system tokens like [/INST] and employing demo-level random search from a collected demo pool. These simple techniques result in surprisingly effective jailbreaking against aligned LLMs (even with advanced defenses). For examples, our method achieves >80% (mostly >95%) ASRs on Llama-2-7B and Llama-3-8B without multiple restarts, even if the models are enhanced by strong defenses such as perplexity detection and/or SmoothLLM, which is challenging for suffix-based jailbreaking. In addition, we conduct comprehensive and elaborate (e.g., making sure to use correct system prompts) evaluations against other aligned LLMs and advanced defenses, where our method consistently achieves nearly 100% ASRs. Our code is available at https://github.com/sail-sg/I-FSJ.
Abstract:The primary challenge in Federated Learning (FL) is to model non-IID distributions across clients, whose fine-grained structure is important to improve knowledge sharing. For example, some knowledge is globally shared across all clients, some is only transferable within a subgroup of clients, and some are client-specific. To capture and exploit this structure, we train models organized in a multi-level structure, called ``Multi-level Additive Models (MAM)'', for better knowledge-sharing across heterogeneous clients and their personalization. In federated MAM (FeMAM), each client is assigned to at most one model per level and its personalized prediction sums up the outputs of models assigned to it across all levels. For the top level, FeMAM trains one global model shared by all clients as FedAvg. For every mid-level, it learns multiple models each assigned to a subgroup of clients, as clustered FL. Every bottom-level model is trained for one client only. In the training objective, each model aims to minimize the residual of the additive predictions by the other models assigned to each client. To approximate the arbitrary structure of non-IID across clients, FeMAM introduces more flexibility and adaptivity to FL by incrementally adding new models to the prediction of each client and reassigning another if necessary, automatically optimizing the knowledge-sharing structure. Extensive experiments show that FeMAM surpasses existing clustered FL and personalized FL methods in various non-IID settings. Our code is available at https://github.com/shutong043/FeMAM.
Abstract:This paper demonstrates that pre-trained language models (PLMs) are strong foundation models for on-device meteorological variables modeling. We present LM-Weather, a generic approach to taming PLMs, that have learned massive sequential knowledge from the universe of natural language databases, to acquire an immediate capability to obtain highly customized models for heterogeneous meteorological data on devices while keeping high efficiency. Concretely, we introduce a lightweight personalized adapter into PLMs and endows it with weather pattern awareness. During communication between clients and the server, low-rank-based transmission is performed to effectively fuse the global knowledge among devices while maintaining high communication efficiency and ensuring privacy. Experiments on real-wold dataset show that LM-Weather outperforms the state-of-the-art results by a large margin across various tasks (e.g., forecasting and imputation at different scales). We provide extensive and in-depth analyses experiments, which verify that LM-Weather can (1) indeed leverage sequential knowledge from natural language to accurately handle meteorological sequence, (2) allows each devices obtain highly customized models under significant heterogeneity, and (3) generalize under data-limited and out-of-distribution (OOD) scenarios.
Abstract:The recent success of large language models (LLMs) has attracted widespread interest to develop role-playing conversational agents personalized to the characteristics and styles of different speakers to enhance their abilities to perform both general and special purpose dialogue tasks. However, the ability to personalize the generated utterances to speakers, whether conducted by human or LLM, has not been well studied. To bridge this gap, our study introduces a novel evaluation challenge: speaker verification in agent-generated conversations, which aimed to verify whether two sets of utterances originate from the same speaker. To this end, we assemble a large dataset collection encompassing thousands of speakers and their utterances. We also develop and evaluate speaker verification models under experiment setups. We further utilize the speaker verification models to evaluate the personalization abilities of LLM-based role-playing models. Comprehensive experiments suggest that the current role-playing models fail in accurately mimicking speakers, primarily due to their inherent linguistic characteristics.
Abstract:In sequential decision-making problems involving sensitive attributes like race and gender, reinforcement learning (RL) agents must carefully consider long-term fairness while maximizing returns. Recent works have proposed many different types of fairness notions, but how unfairness arises in RL problems remains unclear. In this paper, we address this gap in the literature by investigating the sources of inequality through a causal lens. We first analyse the causal relationships governing the data generation process and decompose the effect of sensitive attributes on long-term well-being into distinct components. We then introduce a novel notion called dynamics fairness, which explicitly captures the inequality stemming from environmental dynamics, distinguishing it from those induced by decision-making or inherited from the past. This notion requires evaluating the expected changes in the next state and the reward induced by changing the value of the sensitive attribute while holding everything else constant. To quantitatively evaluate this counterfactual concept, we derive identification formulas that allow us to obtain reliable estimations from data. Extensive experiments demonstrate the effectiveness of the proposed techniques in explaining, detecting, and reducing inequality in reinforcement learning.
Abstract:Automated generation of feedback on programming assignments holds significant benefits for programming education, especially when it comes to advanced assignments. Automated Program Repair techniques, especially Large Language Model based approaches, have gained notable recognition for their potential to fix introductory assignments. However, the programs used for evaluation are relatively simple. It remains unclear how existing approaches perform in repairing programs from higher-level programming courses. To address these limitations, we curate a new advanced student assignment dataset named Defects4DS from a higher-level programming course. Subsequently, we identify the challenges related to fixing bugs in advanced assignments. Based on the analysis, we develop a framework called PaR that is powered by the LLM. PaR works in three phases: Peer Solution Selection, Multi-Source Prompt Generation, and Program Repair. Peer Solution Selection identifies the closely related peer programs based on lexical, semantic, and syntactic criteria. Then Multi-Source Prompt Generation adeptly combines multiple sources of information to create a comprehensive and informative prompt for the last Program Repair stage. The evaluation on Defects4DS and another well-investigated ITSP dataset reveals that PaR achieves a new state-of-the-art performance, demonstrating impressive improvements of 19.94% and 15.2% in repair rate compared to prior state-of-the-art LLM- and symbolic-based approaches, respectively
Abstract:Recently, foundation models, particularly large language models (LLMs), have demonstrated an impressive ability to adapt to various tasks by fine-tuning large amounts of instruction data. Notably, federated foundation models emerge as a privacy preservation method to fine-tune models collaboratively under federated learning (FL) settings by leveraging many distributed datasets with non-IID data. To alleviate communication and computation overhead, parameter-efficient methods are introduced for efficiency, and some research adapted personalization methods to federated foundation models for better user preferences alignment. However, a critical gap in existing research is the neglect of test-time distribution shifts in real-world applications. Therefore, to bridge this gap, we propose a new setting, termed test-time personalization, which not only concentrates on the targeted local task but also extends to other tasks that exhibit test-time distribution shifts. To address challenges in this new setting, we explore a simple yet effective solution to learn a comprehensive foundation model. Specifically, a dual-personalizing adapter architecture (FedDPA) is proposed, comprising a global adapter and a local adapter for addressing test-time distribution shifts and personalization, respectively. Additionally, we introduce an instance-wise dynamic weighting mechanism to optimize the balance between the global and local adapters, enhancing overall performance. The effectiveness of the proposed method has been evaluated on benchmark datasets across different NLP tasks.
Abstract:In this paper, we address the challenge of detecting hateful memes in the low-resource setting where only a few labeled examples are available. Our approach leverages the compositionality of Low-rank adaptation (LoRA), a widely used parameter-efficient tuning technique. We commence by fine-tuning large language models (LLMs) with LoRA on selected tasks pertinent to hateful meme detection, thereby generating a suite of LoRA modules. These modules are capable of essential reasoning skills for hateful meme detection. We then use the few available annotated samples to train a module composer, which assigns weights to the LoRA modules based on their relevance. The model's learnable parameters are directly proportional to the number of LoRA modules. This modularized network, underpinned by LLMs and augmented with LoRA modules, exhibits enhanced generalization in the context of hateful meme detection. Our evaluation spans three datasets designed for hateful meme detection in a few-shot learning context. The proposed method demonstrates superior performance to traditional in-context learning, which is also more computationally intensive during inference.We then use the few available annotated samples to train a module composer, which assigns weights to the LoRA modules based on their relevance. The model's learnable parameters are directly proportional to the number of LoRA modules. This modularized network, underpinned by LLMs and augmented with LoRA modules, exhibits enhanced generalization in the context of hateful meme detection. Our evaluation spans three datasets designed for hateful meme detection in a few-shot learning context. The proposed method demonstrates superior performance to traditional in-context learning, which is also more computationally intensive during inference.