Large language models (LLMs) are often fine-tuned for use on downstream tasks, though this can degrade capabilities learned during previous training. This phenomenon, often referred to as catastrophic forgetting, has important potential implications for the safety of deployed models. In this work, we first show that models trained on downstream tasks forget their safety tuning to a greater extent than models trained in the opposite order. Second, we show that forgetting disproportionately impacts safety information about certain groups. To quantify this phenomenon, we define a new metric we term biased forgetting. We conduct a systematic evaluation of the effects of task ordering on forgetting and apply mitigations that can help the model recover from the forgetting observed. We hope our findings can better inform methods for chaining the finetuning of LLMs in continual learning settings to enable training of safer and less toxic models.

One of the most challenging problems facing NLP today is evaluation. Some of the most pressing issues pertain to benchmark saturation, data contamination, and diversity in the quality of test examples. To address these concerns, we propose Selection Methodology for Accurate, Reduced, and Targeted (SMART) filtering, a novel approach to select a high-quality subset of examples from existing benchmark datasets by systematically removing less informative and less challenging examples. Our approach applies three filtering criteria, removing (i) easy examples, (ii) data-contaminated examples, and (iii) examples that are similar to each other based on distance in an embedding space. We demonstrate the effectiveness of SMART on three multiple choice QA datasets, where our methodology increases efficiency by reducing dataset size by 48\% on average, while increasing Pearson correlation with rankings from ChatBot Arena, a more open-ended human evaluation setting. Our method enables us to be more efficient, whether using SMART to make new benchmarks more challenging or to revitalize older datasets, while still preserving the relative model rankings.

For can affect multiple choice tests, for example, example, NLP model accuracy has been shown to when answers are presented in a different order be fairly brittle. For example, accuracy can drop during retest (Krosnick and Fabrigar, 1991; when researchers apply input alterations based Tellinghuisen and Sulikowski, 2008; Lions et al., on paraphrasing (Gan and Ng, 2019), word order 2022). However, as models do not have the biological changes (Gauthier and Levy, 2019; Ribeiro et al., limitations of humans, we may expect them 2020; Sinha et al., 2021a, 2022; Allen-Zhu and Li, to exhibit less variation than humans, or possibly 2023a,b; Berglund et al., 2023; Golovneva et al., even none at all. Thus, we claim that a model 2024; Kitouni et al., 2024), or other minor, largely should be robust to answer order changes: if it gets meaning-preserving input variations or perturbations the correct answer to a question when the answer (Belinkov and Bisk, 2018; Ebrahimi et al., is labeled'A', it should also always get the correct 2018; Jiang et al., 2020; Gao et al., 2021; Li et al., answer when it is labeled'C'. Put another way, 2021; Sinha et al., 2021b; Moradi and Samwald, the model should select the same answer for each 2021; Papakipos and Bitton, 2022; Qian et al., question, regardless of its label, for every possible 2022; Goodarzi et al., 2023; Sinha et al., 2023).

As generative large language models (LLMs) grow more performant and prevalent, we must develop comprehensive enough tools to measure and improve their fairness. Different prompt-based datasets can be used to measure social bias across multiple text domains and demographic axes, meaning that testing LLMs on more datasets can potentially help us characterize their biases more fully, and better ensure equal and equitable treatment of marginalized demographic groups. In this work, our focus is two-fold: (1) Benchmarking: a comparison of 6 different prompt-based bias and toxicity metrics across 12 demographic axes and 5 families of generative LLMs. Out of those 6 metrics, AdvPromptSet and HolisticBiasR are novel datasets proposed in the paper. The comparison of those benchmarks gives us insights about the bias and toxicity of the compared models. Therefore, we explore the frequency of demographic terms in common LLM pre-training corpora and how this may relate to model biases. (2) Mitigation: we conduct a comprehensive study of how well 3 bias/toxicity mitigation techniques perform across our suite of measurements. ROBBIE aims to provide insights for practitioners while deploying a model, emphasizing the need to not only measure potential harms, but also understand how they arise by characterizing the data, mitigate harms once found, and balance any trade-offs. We open-source our analysis code in hopes of encouraging broader measurements of bias in future LLMs.

Many cognitive approaches to well-being, such as recognizing and reframing unhelpful thoughts, have received considerable empirical support over the past decades, yet still lack truly widespread adoption in self-help format. A barrier to that adoption is a lack of adequately specific and diverse dedicated practice material. This work examines whether current language models can be leveraged to both produce a virtually unlimited quantity of practice material illustrating standard unhelpful thought patterns matching specific given contexts, and generate suitable positive reframing proposals. We propose PATTERNREFRAME, a novel dataset of about 10k examples of thoughts containing unhelpful thought patterns conditioned on a given persona, accompanied by about 27k positive reframes. By using this dataset to train and/or evaluate current models, we show that existing models can already be powerful tools to help generate an abundance of tailored practice material and hypotheses, with no or minimal additional model training required.

We present BlenderBot 3x, an update on the conversational model BlenderBot 3, which is now trained using organic conversation and feedback data from participating users of the system in order to improve both its skills and safety. We are publicly releasing the participating de-identified interaction data for use by the research community, in order to spur further progress. Training models with organic data is challenging because interactions with people "in the wild" include both high quality conversations and feedback, as well as adversarial and toxic behavior. We study techniques that enable learning from helpful teachers while avoiding learning from people who are trying to trick the model into unhelpful or toxic responses. BlenderBot 3x is both preferred in conversation to BlenderBot 3, and is shown to produce safer responses in challenging situations. While our current models are still far from perfect, we believe further improvement can be achieved by continued use of the techniques explored in this work.

Current open-domain conversational models can easily be made to talk in inadequate ways. Online learning from conversational feedback given by the conversation partner is a promising avenue for a model to improve and adapt, so as to generate fewer of these safety failures. However, current state-of-the-art models tend to react to feedback with defensive or oblivious responses. This makes for an unpleasant experience and may discourage conversation partners from giving feedback in the future. This work proposes SaFeRDialogues, a task and dataset of graceful responses to conversational feedback about safety failures. We collect a dataset of 10k dialogues demonstrating safety failures, feedback signaling them, and a response acknowledging the feedback. We show how fine-tuning on this dataset results in conversations that human raters deem considerably more likely to lead to a civil conversation, without sacrificing engagingness or general conversational ability.