InstructRetro: Instruction Tuning post Retrieval-Augmented Pretraining

Pretraining auto-regressive large language models (LLMs) with retrieval demonstrates better perplexity and factual accuracy by leveraging external databases. However, the size of existing pretrained retrieval-augmented LLM is still limited (e.g., Retro has 7.5B parameters), which limits the effectiveness of instruction tuning and zero-shot generalization. In this work, we introduce Retro 48B, the largest LLM pretrained with retrieval before instruction tuning. Specifically, we continue to pretrain the 43B GPT model on additional 100 billion tokens using the Retro augmentation method by retrieving from 1.2 trillion tokens. The obtained foundation model, Retro 48B, largely outperforms the original 43B GPT in terms of perplexity. After instruction tuning on Retro, InstructRetro demonstrates significant improvement over the instruction tuned GPT on zero-shot question answering (QA) tasks. Specifically, the average improvement of InstructRetro is 7% over its GPT counterpart across 8 short-form QA tasks, and 10% over GPT across 4 challenging long-form QA tasks. Surprisingly, we find that one can ablate the encoder from InstructRetro architecture and directly use its decoder backbone, while achieving comparable results. We hypothesize that pretraining with retrieval makes its decoder good at incorporating context for QA. Our results highlights the promising direction to obtain a better GPT decoder for QA through continued pretraining with retrieval before instruction tuning.

DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models

Generative Pre-trained Transformer (GPT) models have exhibited exciting progress in capabilities, capturing the interest of practitioners and the public alike. Yet, while the literature on the trustworthiness of GPT models remains limited, practitioners have proposed employing capable GPT models for sensitive applications to healthcare and finance - where mistakes can be costly. To this end, this work proposes a comprehensive trustworthiness evaluation for large language models with a focus on GPT-4 and GPT-3.5, considering diverse perspectives - including toxicity, stereotype bias, adversarial robustness, out-of-distribution robustness, robustness on adversarial demonstrations, privacy, machine ethics, and fairness. Based on our evaluations, we discover previously unpublished vulnerabilities to trustworthiness threats. For instance, we find that GPT models can be easily misled to generate toxic and biased outputs and leak private information in both training data and conversation history. We also find that although GPT-4 is usually more trustworthy than GPT-3.5 on standard benchmarks, GPT-4 is more vulnerable given jailbreaking system or user prompts, potentially due to the reason that GPT-4 follows the (misleading) instructions more precisely. Our work illustrates a comprehensive trustworthiness evaluation of GPT models and sheds light on the trustworthiness gaps. Our benchmark is publicly available at https://decodingtrust.github.io/.

Can Public Large Language Models Help Private Cross-device Federated Learning?

We study (differentially) private federated learning (FL) of language models. The language models in cross-device FL are relatively small, which can be trained with meaningful formal user-level differential privacy (DP) guarantees when massive parallelism in training is enabled by the participation of a moderate size of users. Recently, public data has been used to improve privacy-utility trade-offs for both large and small language models. In this work, we provide a systematic study of using large-scale public data and LLMs to help differentially private training of on-device FL models, and further improve the privacy-utility tradeoff by techniques of distillation. Moreover, we propose a novel distribution matching algorithm with theoretical grounding to sample public data close to private data distribution, which significantly improves the sample efficiency of (pre-)training on public data. The proposed method is efficient and effective for training private model by taking advantage of public data, especially for customized on-device architectures that do not have ready-to-use pre-trained models.

Shall We Pretrain Autoregressive Language Models with Retrieval? A Comprehensive Study

Large decoder-only language models (LMs) can be largely improved in terms of perplexity by retrieval (e.g., RETRO), but its impact on text generation quality and downstream task accuracy is unclear. Thus, it is still an open question: shall we pretrain large autoregressive LMs with retrieval? To answer it, we perform a comprehensive study on a scalable pre-trained retrieval-augmented LM (i.e., RETRO) compared with standard GPT and retrieval-augmented GPT incorporated at fine-tuning or inference stages. We first provide the recipe to reproduce RETRO up to 9.5B parameters while retrieving a text corpus with 330B tokens. Based on that, we have the following novel findings: i) RETRO outperforms GPT on text generation with much less degeneration (i.e., repetition), moderately higher factual accuracy, and slightly lower toxicity with a nontoxic retrieval database. ii) On the LM Evaluation Harness benchmark, RETRO largely outperforms GPT on knowledge-intensive tasks, but is on par with GPT on other tasks. Furthermore, we introduce a simple variant of the model, RETRO++, which largely improves open-domain QA results of original RETRO (e.g., EM score +8.6 on Natural Question) and significantly outperforms retrieval-augmented GPT across different model sizes. Our findings highlight the promising direction of pretraining autoregressive LMs with retrieval as future foundation models. We release our implementation at: https://github.com/NVIDIA/Megatron-LM#retro

FOCUS: Fairness via Agent-Awareness for Federated Learning on Heterogeneous Data

Federated learning (FL) provides an effective paradigm to train machine learning models over distributed data with privacy protection. However, recent studies show that FL is subject to various security, privacy, and fairness threats due to the potentially malicious and heterogeneous local agents. For instance, it is vulnerable to local adversarial agents who only contribute low-quality data, with the goal of harming the performance of those with high-quality data. This kind of attack hence breaks existing definitions of fairness in FL that mainly focus on a certain notion of performance parity. In this work, we aim to address this limitation and propose a formal definition of fairness via agent-awareness for FL (FAA), which takes the heterogeneous data contributions of local agents into account. In addition, we propose a fair FL training algorithm based on agent clustering (FOCUS) to achieve FAA. Theoretically, we prove the convergence and optimality of FOCUS under mild conditions for linear models and general convex loss functions with bounded smoothness. We also prove that FOCUS always achieves higher fairness measured by FAA compared with standard FedAvg protocol under both linear models and general convex loss functions. Empirically, we evaluate FOCUS on four datasets, including synthetic data, images, and texts under different settings, and we show that FOCUS achieves significantly higher fairness based on FAA while maintaining similar or even higher prediction accuracy compared with FedAvg.

SemAttack: Natural Textual Attacks via Different Semantic Spaces

Recent studies show that pre-trained language models (LMs) are vulnerable to textual adversarial attacks. However, existing attack methods either suffer from low attack success rates or fail to search efficiently in the exponentially large perturbation space. We propose an efficient and effective framework SemAttack to generate natural adversarial text by constructing different semantic perturbation functions. In particular, SemAttack optimizes the generated perturbations constrained on generic semantic spaces, including typo space, knowledge space (e.g., WordNet), contextualized semantic space (e.g., the embedding space of BERT clusterings), or the combination of these spaces. Thus, the generated adversarial texts are more semantically close to the original inputs. Extensive experiments reveal that state-of-the-art (SOTA) large-scale LMs (e.g., DeBERTa-v2) and defense strategies (e.g., FreeLB) are still vulnerable to SemAttack. We further demonstrate that SemAttack is general and able to generate natural adversarial texts for different languages (e.g., English and Chinese) with high attack success rates. Human evaluations also confirm that our generated adversarial texts are natural and barely affect human performance. Our code is publicly available at https://github.com/AI-secure/SemAttack.

Exploring the Limits of Domain-Adaptive Training for Detoxifying Large-Scale Language Models

Pre-trained language models (LMs) are shown to easily generate toxic language. In this work, we systematically explore domain-adaptive training to reduce the toxicity of language models. We conduct this study on three dimensions: training corpus, model size, and parameter efficiency. For the training corpus, we propose to leverage the generative power of LMs and generate nontoxic datasets for domain-adaptive training, which mitigates the exposure bias and is shown to be more data-efficient than using a curated pre-training corpus. We demonstrate that the self-generation method consistently outperforms the existing baselines across various model sizes on both automatic and human evaluations, even when it uses a 1/3 smaller training corpus. We then comprehensively study detoxifying LMs with parameter sizes ranging from 126M up to 530B (3x larger than GPT-3), a scale that has never been studied before. We find that i) large LMs have similar toxicity levels as smaller ones given the same pre-training corpus, and ii) large LMs require more endeavor to detoxify. We also explore parameter-efficient training methods for detoxification. We demonstrate that adding and training adapter-only layers in LMs not only saves a lot of parameters but also achieves a better trade-off between toxicity and perplexity than whole model adaptation for the large-scale models.

Certifying Out-of-Domain Generalization for Blackbox Functions

Certifying the robustness of model performance under bounded data distribution shifts has recently attracted intensive interests under the umbrella of distributional robustness. However, existing techniques either make strong assumptions on the model class and loss functions that can be certified, such as smoothness expressed via Lipschitz continuity of gradients, or require to solve complex optimization problems. As a result, the wider application of these techniques is currently limited by its scalability and flexibility -- these techniques often do not scale to large-scale datasets with modern deep neural networks or cannot handle loss functions which may be non-smooth, such as the 0-1 loss. In this paper, we focus on the problem of certifying distributional robustness for black box models and bounded losses, without other assumptions. We propose a novel certification framework given bounded distance of mean and variance of two distributions. Our certification technique scales to ImageNet-scale datasets, complex models, and a diverse range of loss functions. We then focus on one specific application enabled by such scalability and flexibility, i.e., certifying out-of-domain generalization for large neural networks and loss functions such as accuracy and AUC. We experimentally validate our certification method on a number of datasets, ranging from ImageNet, where we provide the first non-vacuous certified out-of-domain generalization, to smaller classification tasks where we are able to compare with the state-of-the-art and show that our method performs considerably better.

Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models

Large-scale pre-trained language models have achieved tremendous success across a wide range of natural language understanding (NLU) tasks, even surpassing human performance. However, recent studies reveal that the robustness of these models can be challenged by carefully crafted textual adversarial examples. While several individual datasets have been proposed to evaluate model robustness, a principled and comprehensive benchmark is still missing. In this paper, we present Adversarial GLUE (AdvGLUE), a new multi-task benchmark to quantitatively and thoroughly explore and evaluate the vulnerabilities of modern large-scale language models under various types of adversarial attacks. In particular, we systematically apply 14 textual adversarial attack methods to GLUE tasks to construct AdvGLUE, which is further validated by humans for reliable annotations. Our findings are summarized as follows. (i) Most existing adversarial attack algorithms are prone to generating invalid or ambiguous adversarial examples, with around 90% of them either changing the original semantic meanings or misleading human annotators as well. Therefore, we perform a careful filtering process to curate a high-quality benchmark. (ii) All the language models and robust training methods we tested perform poorly on AdvGLUE, with scores lagging far behind the benign accuracy. We hope our work will motivate the development of new adversarial attacks that are more stealthy and semantic-preserving, as well as new robust language models against sophisticated adversarial attacks. AdvGLUE is available at https://adversarialglue.github.io.