LifeTox: Unveiling Implicit Toxicity in Life Advice

As large language models become increasingly integrated into daily life, detecting implicit toxicity across diverse contexts is crucial. To this end, we introduce LifeTox, a dataset designed for identifying implicit toxicity within a broad range of advice-seeking scenarios. Unlike existing safety datasets, LifeTox comprises diverse contexts derived from personal experiences through open-ended questions. Experiments demonstrate that RoBERTa fine-tuned on LifeTox matches or surpasses the zero-shot performance of large language models in toxicity classification tasks. These results underscore the efficacy of LifeTox in addressing the complex challenges inherent in implicit toxicity.

Prometheus: Inducing Fine-grained Evaluation Capability in Language Models

Recently, using a powerful proprietary Large Language Model (LLM) (e.g., GPT-4) as an evaluator for long-form responses has become the de facto standard. However, for practitioners with large-scale evaluation tasks and custom criteria in consideration (e.g., child-readability), using proprietary LLMs as an evaluator is unreliable due to the closed-source nature, uncontrolled versioning, and prohibitive costs. In this work, we propose Prometheus, a fully open-source LLM that is on par with GPT-4's evaluation capabilities when the appropriate reference materials (reference answer, score rubric) are accompanied. We first construct the Feedback Collection, a new dataset that consists of 1K fine-grained score rubrics, 20K instructions, and 100K responses and language feedback generated by GPT-4. Using the Feedback Collection, we train Prometheus, a 13B evaluator LLM that can assess any given long-form text based on customized score rubric provided by the user. Experimental results show that Prometheus scores a Pearson correlation of 0.897 with human evaluators when evaluating with 45 customized score rubrics, which is on par with GPT-4 (0.882), and greatly outperforms ChatGPT (0.392). Furthermore, measuring correlation with GPT-4 with 1222 customized score rubrics across four benchmarks (MT Bench, Vicuna Bench, Feedback Bench, Flask Eval) shows similar trends, bolstering Prometheus's capability as an evaluator LLM. Lastly, Prometheus achieves the highest accuracy on two human preference benchmarks (HHH Alignment & MT Bench Human Judgment) compared to open-sourced reward models explicitly trained on human preference datasets, highlighting its potential as an universal reward model. We open-source our code, dataset, and model at https://github.com/kaistAI/Prometheus.

KoBBQ: Korean Bias Benchmark for Question Answering

The BBQ (Bias Benchmark for Question Answering) dataset enables the evaluation of the social biases that language models (LMs) exhibit in downstream tasks. However, it is challenging to adapt BBQ to languages other than English as social biases are culturally dependent. In this paper, we devise a process to construct a non-English bias benchmark dataset by leveraging the English BBQ dataset in a culturally adaptive way and present the KoBBQ dataset for evaluating biases in Question Answering (QA) tasks in Korean. We identify samples from BBQ into three classes: Simply-Translated (can be used directly after cultural translation), Target-Modified (requires localization in target groups), and Sample-Removed (does not fit Korean culture). We further enhance the cultural relevance to Korean culture by adding four new categories of bias specific to Korean culture and newly creating samples based on Korean literature. KoBBQ consists of 246 templates and 4,740 samples across 12 categories of social bias. Using KoBBQ, we measure the accuracy and bias scores of several state-of-the-art multilingual LMs. We demonstrate the differences in the bias of LMs in Korean and English, clarifying the need for hand-crafted data considering cultural differences.

ProPILE: Probing Privacy Leakage in Large Language Models

The rapid advancement and widespread use of large language models (LLMs) have raised significant concerns regarding the potential leakage of personally identifiable information (PII). These models are often trained on vast quantities of web-collected data, which may inadvertently include sensitive personal data. This paper presents ProPILE, a novel probing tool designed to empower data subjects, or the owners of the PII, with awareness of potential PII leakage in LLM-based services. ProPILE lets data subjects formulate prompts based on their own PII to evaluate the level of privacy intrusion in LLMs. We demonstrate its application on the OPT-1.3B model trained on the publicly available Pile dataset. We show how hypothetical data subjects may assess the likelihood of their PII being included in the Pile dataset being revealed. ProPILE can also be leveraged by LLM service providers to effectively evaluate their own levels of PII leakage with more powerful prompts specifically tuned for their in-house models. This tool represents a pioneering step towards empowering the data subjects for their awareness and control over their own data on the web.

KoSBi: A Dataset for Mitigating Social Bias Risks Towards Safer Large Language Model Application

Large language models (LLMs) learn not only natural text generation abilities but also social biases against different demographic groups from real-world data. This poses a critical risk when deploying LLM-based applications. Existing research and resources are not readily applicable in South Korea due to the differences in language and culture, both of which significantly affect the biases and targeted demographic groups. This limitation requires localized social bias datasets to ensure the safe and effective deployment of LLMs. To this end, we present KO SB I, a new social bias dataset of 34k pairs of contexts and sentences in Korean covering 72 demographic groups in 15 categories. We find that through filtering-based moderation, social biases in generated content can be reduced by 16.47%p on average for HyperCLOVA (30B and 82B), and GPT-3.

SQuARe: A Large-Scale Dataset of Sensitive Questions and Acceptable Responses Created Through Human-Machine Collaboration

The potential social harms that large language models pose, such as generating offensive content and reinforcing biases, are steeply rising. Existing works focus on coping with this concern while interacting with ill-intentioned users, such as those who explicitly make hate speech or elicit harmful responses. However, discussions on sensitive issues can become toxic even if the users are well-intentioned. For safer models in such scenarios, we present the Sensitive Questions and Acceptable Response (SQuARe) dataset, a large-scale Korean dataset of 49k sensitive questions with 42k acceptable and 46k non-acceptable responses. The dataset was constructed leveraging HyperCLOVA in a human-in-the-loop manner based on real news headlines. Experiments show that acceptable response generation significantly improves for HyperCLOVA and GPT-3, demonstrating the efficacy of this dataset.

Query-Efficient Black-Box Red Teaming via Bayesian Optimization

The deployment of large-scale generative models is often restricted by their potential risk of causing harm to users in unpredictable ways. We focus on the problem of black-box red teaming, where a red team generates test cases and interacts with the victim model to discover a diverse set of failures with limited query access. Existing red teaming methods construct test cases based on human supervision or language model (LM) and query all test cases in a brute-force manner without incorporating any information from past evaluations, resulting in a prohibitively large number of queries. To this end, we propose Bayesian red teaming (BRT), novel query-efficient black-box red teaming methods based on Bayesian optimization, which iteratively identify diverse positive test cases leading to model failures by utilizing the pre-defined user input pool and the past evaluations. Experimental results on various user input pools demonstrate that our method consistently finds a significantly larger number of diverse positive test cases under the limited query budget than the baseline methods. The source code is available at https://github.com/snu-mllab/Bayesian-Red-Teaming.

Who Wrote this Code? Watermarking for Code Generation

Large language models for code have recently shown remarkable performance in generating executable code. However, this rapid advancement has been accompanied by many legal and ethical concerns, such as code licensing issues, code plagiarism, and malware generation, making watermarking machine-generated code a very timely problem. Despite such imminent needs, we discover that existing watermarking and machine-generated text detection methods for LLMs fail to function with code generation tasks properly. Hence, in this work, we propose a new watermarking method, SWEET, that significantly improves upon previous approaches when watermarking machine-generated code. Our proposed method selectively applies watermarking to the tokens with high enough entropy, surpassing a defined threshold. The experiments on code generation benchmarks show that our watermarked code has superior quality compared to code produced by the previous state-of-the-art LLM watermarking method. Furthermore, our watermark method also outperforms DetectGPT for the task of machine-generated code detection.

Critic-Guided Decoding for Controlled Text Generation

Steering language generation towards objectives or away from undesired content has been a long-standing goal in utilizing language models (LM). Recent work has demonstrated reinforcement learning and weighted decoding as effective approaches to achieve a higher level of language control and quality with pros and cons. In this work, we propose a novel critic decoding method for controlled language generation (CriticControl) that combines the strengths of reinforcement learning and weighted decoding. Specifically, we adopt the actor-critic framework to train an LM-steering critic from non-differentiable reward models. And similar to weighted decoding, our method freezes the language model and manipulates the output token distribution using called critic, improving training efficiency and stability. Evaluation of our method on three controlled generation tasks, namely topic control, sentiment control, and detoxification, shows that our approach generates more coherent and well-controlled texts than previous methods. In addition, CriticControl demonstrates superior generalization ability in zero-shot settings. Human evaluation studies also corroborate our findings.

Beyond Fact Verification: Comparing and Contrasting Claims on Contentious Topics

As the importance of identifying misinformation is increasing, many researchers focus on verifying textual claims on the web. One of the most popular tasks to achieve this is fact verification, which retrieves an evidence sentence from a large knowledge source such as Wikipedia to either verify or refute each factual claim. However, while such problem formulation is helpful for detecting false claims and fake news, it is not applicable to catching subtle differences in factually consistent claims which still might implicitly bias the readers, especially in contentious topics such as political, gender, or racial issues. In this study, we propose ClaimDiff, a novel dataset to compare the nuance between claim pairs in both a discriminative and a generative manner, with the underlying assumption that one is not necessarily more true than the other. This differs from existing fact verification datasets that verify the target sentence with respect to an absolute truth. We hope this task assists people in making more informed decisions among various sources of media.