PISTOL: Dataset Compilation Pipeline for Structural Unlearning of LLMs

Recently, machine unlearning, which seeks to erase specific data stored in the pre-trained or fine-tuned models, has emerged as a crucial protective measure for LLMs. However, unlearning approaches for LLMs that have been considered thus far have focused on the removal of independent data points and have not taken into account that the stored facts are logically connected to one another and form an implicit knowledge graph. To facilitate the development of structural unlearning methods, which are essential for the practical application of unlearning, we propose PISTOL, a pipeline for compiling multi-scenario datasets for benchmarking structural LLM unlearning. Additionally, leveraging sample datasets synthesized using PISTOL, we conducted benchmarks with four distinct unlearning methods on both Llama2-7B and Mistral-7B models. This analysis helps to illustrate the prevailing challenges in effectively and robustly removing highly inter-connected data, batched data, or data skewed towards a specific domain. It also highlights the choice of pre-trained model can impact unlearning performance. This work not only advances our understandings on the limitation of current LLMs unlearning methods and proposes future research directions, but also provides a replicable framework for ongoing exploration and validation in the field.

Quantifying Generative Media Bias with a Corpus of Real-world and Generated News Articles

Large language models (LLMs) are increasingly being utilised across a range of tasks and domains, with a burgeoning interest in their application within the field of journalism. This trend raises concerns due to our limited understanding of LLM behaviour in this domain, especially with respect to political bias. Existing studies predominantly focus on LLMs undertaking political questionnaires, which offers only limited insights into their biases and operational nuances. To address this gap, our study establishes a new curated dataset that contains 2,100 human-written articles and utilises their descriptions to generate 56,700 synthetic articles using nine LLMs. This enables us to analyse shifts in properties between human-authored and machine-generated articles, with this study focusing on political bias, detecting it using both supervised models and LLMs. Our findings reveal significant disparities between base and instruction-tuned LLMs, with instruction-tuned models exhibiting consistent political bias. Furthermore, we are able to study how LLMs behave as classifiers, observing their display of political bias even in this role. Overall, for the first time within the journalistic domain, this study outlines a framework and provides a structured dataset for quantifiable experiments, serving as a foundation for further research into LLM political bias and its implications.

Quantifying Variance in Evaluation Benchmarks

Evaluation benchmarks are the cornerstone of measuring capabilities of large language models (LLMs), as well as driving progress in said capabilities. Originally designed to make claims about capabilities (or lack thereof) in fully pretrained models, evaluation benchmarks are now also extensively used to decide between various training choices. Despite this widespread usage, we rarely quantify the variance in our evaluation benchmarks, which dictates whether differences in performance are meaningful. Here, we define and measure a range of metrics geared towards measuring variance in evaluation benchmarks, including seed variance across initialisations, and monotonicity during training. By studying a large number of models -- both openly available and pretrained from scratch -- we provide empirical estimates for a variety of variance metrics, with considerations and recommendations for practitioners. We also evaluate the utility and tradeoffs of continuous versus discrete performance measures and explore options for better understanding and reducing this variance. We find that simple changes, such as framing choice tasks (like MMLU) as completion tasks, can often reduce variance for smaller scale ($\sim$7B) models, while more involved methods inspired from human testing literature (such as item analysis and item response theory) struggle to meaningfully reduce variance. Overall, our work provides insights into variance in evaluation benchmarks, suggests LM-specific techniques to reduce variance, and more generally encourages practitioners to carefully factor in variance when comparing models.

Words Worth a Thousand Pictures: Measuring and Understanding Perceptual Variability in Text-to-Image Generation

Diffusion models are the state of the art in text-to-image generation, but their perceptual variability remains understudied. In this paper, we examine how prompts affect image variability in black-box diffusion-based models. We propose W1KP, a human-calibrated measure of variability in a set of images, bootstrapped from existing image-pair perceptual distances. Current datasets do not cover recent diffusion models, thus we curate three test sets for evaluation. Our best perceptual distance outperforms nine baselines by up to 18 points in accuracy, and our calibration matches graded human judgements 78% of the time. Using W1KP, we study prompt reusability and show that Imagen prompts can be reused for 10-50 random seeds before new images become too similar to already generated images, while Stable Diffusion XL and DALL-E 3 can be reused 50-200 times. Lastly, we analyze 56 linguistic features of real prompts, finding that the prompt's length, CLIP embedding norm, concreteness, and word senses influence variability most. As far as we are aware, we are the first to analyze diffusion variability from a visuolinguistic perspective. Our project page is at http://w1kp.com

IrokoBench: A New Benchmark for African Languages in the Age of Large Language Models

Despite the widespread adoption of Large language models (LLMs), their remarkable capabilities remain limited to a few high-resource languages. Additionally, many low-resource languages (e.g. African languages) are often evaluated only on basic text classification tasks due to the lack of appropriate or comprehensive benchmarks outside of high-resource languages. In this paper, we introduce IrokoBench -- a human-translated benchmark dataset for 16 typologically-diverse low-resource African languages covering three tasks: natural language inference~(AfriXNLI), mathematical reasoning~(AfriMGSM), and multi-choice knowledge-based QA~(AfriMMLU). We use IrokoBench to evaluate zero-shot, few-shot, and translate-test settings~(where test sets are translated into English) across 10 open and four proprietary LLMs. Our evaluation reveals a significant performance gap between high-resource languages~(such as English and French) and low-resource African languages. We observe a significant performance gap between open and proprietary models, with the highest performing open model, Aya-101 only at 58\% of the best-performing proprietary model GPT-4o performance. Machine translating the test set to English before evaluation helped to close the gap for larger models that are English-centric, like LLaMa 3 70B. These findings suggest that more efforts are needed to develop and adapt LLMs for African languages.

Transferring Troubles: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning

The implications of backdoor attacks on English-centric large language models (LLMs) have been widely examined - such attacks can be achieved by embedding malicious behaviors during training and activated under specific conditions that trigger malicious outputs. However, the impact of backdoor attacks on multilingual models remains under-explored. Our research focuses on cross-lingual backdoor attacks against multilingual LLMs, particularly investigating how poisoning the instruction-tuning data in one or two languages can affect the outputs in languages whose instruction-tuning data was not poisoned. Despite its simplicity, our empirical analysis reveals that our method exhibits remarkable efficacy in models like mT5, BLOOM, and GPT-3.5-turbo, with high attack success rates, surpassing 95% in several languages across various scenarios. Alarmingly, our findings also indicate that larger models show increased susceptibility to transferable cross-lingual backdoor attacks, which also applies to LLMs predominantly pre-trained on English data, such as Llama2, Llama3, and Gemma. Moreover, our experiments show that triggers can still work even after paraphrasing, and the backdoor mechanism proves highly effective in cross-lingual response settings across 25 languages, achieving an average attack success rate of 50%. Our study aims to highlight the vulnerabilities and significant security risks present in current multilingual LLMs, underscoring the emergent need for targeted security measures.

Prompt Optimisation with Random Sampling

Using the generative nature of a language model to generate task-relevant separators has shown competitive results compared to human-curated prompts like "TL;DR". We demonstrate that even randomly chosen tokens from the vocabulary as separators can achieve near-state-of-the-art performance. We analyse this phenomenon in detail using three different random generation strategies, establishing that the language space is rich with potential good separators, regardless of the underlying language model size. These observations challenge the common assumption that an effective prompt should be human-readable or task-relevant. Experimental results show that using random separators leads to an average 16% relative improvement across nine text classification tasks on seven language models, compared to human-curated separators, and is on par with automatic prompt searching methods.

AfriMTE and AfriCOMET: Empowering COMET to Embrace Under-resourced African Languages

Despite the progress we have recorded in scaling multilingual machine translation (MT) models and evaluation data to several under-resourced African languages, it is difficult to measure accurately the progress we have made on these languages because evaluation is often performed on n-gram matching metrics like BLEU that often have worse correlation with human judgments. Embedding-based metrics such as COMET correlate better; however, lack of evaluation data with human ratings for under-resourced languages, complexity of annotation guidelines like Multidimensional Quality Metrics (MQM), and limited language coverage of multilingual encoders have hampered their applicability to African languages. In this paper, we address these challenges by creating high-quality human evaluation data with a simplified MQM guideline for error-span annotation and direct assessment (DA) scoring for 13 typologically diverse African languages. Furthermore, we develop AfriCOMET, a COMET evaluation metric for African languages by leveraging DA training data from high-resource languages and African-centric multilingual encoder (AfroXLM-Roberta) to create the state-of-the-art evaluation metric for African languages MT with respect to Spearman-rank correlation with human judgments (+0.406).

How good are Large Language Models on African Languages?

Recent advancements in natural language processing have led to the proliferation of large language models (LLMs). These models have been shown to yield good performance, using in-context learning, even on unseen tasks and languages. Additionally, they have been widely adopted as language-model-as-a-service commercial APIs like GPT-4 API. However, their performance on African languages is largely unknown. We present an analysis of three popular large language models (mT0, LLaMa 2, and GPT-4) on five tasks (news topic classification, sentiment classification, machine translation, question answering, and named entity recognition) across 30 African languages, spanning different language families and geographical regions. Our results suggest that all LLMs produce below-par performance on African languages, and there is a large gap in performance compared to high-resource languages like English most tasks. We find that GPT-4 has an average or impressive performance on classification tasks but very poor results on generative tasks like machine translation. Surprisingly, we find that mT0 had the best overall on cross-lingual QA, better than the state-of-the-art supervised model (i.e. fine-tuned mT5) and GPT-4 on African languages. Overall, LLaMa 2 records the worst performance due to its limited multilingual capabilities and English-centric pre-training corpus. In general, our findings present a call-to-action to ensure African languages are well represented in large language models, given their growing popularity.

Using Natural Language Explanations to Improve Robustness of In-context Learning for Natural Language Inference

Recent studies have demonstrated that large language models (LLMs) excel in diverse tasks through in-context learning (ICL) facilitated by task-specific prompts and examples. However, the existing literature shows that ICL encounters performance deterioration when exposed to adversarial inputs. Enhanced performance has been observed when ICL is augmented with natural language explanations (NLEs) (we refer to it as X-ICL). Thus, this work investigates whether X-ICL can improve the robustness of LLMs on a suite of seven adversarial and challenging natural language inference datasets. Moreover, we introduce a new approach to X-ICL by prompting an LLM (ChatGPT in our case) with few human-generated NLEs to produce further NLEs (we call it ChatGPT few-shot), which we show superior to both ChatGPT zero-shot and human-generated NLEs alone. We evaluate five popular LLMs (GPT3.5-turbo, LLaMa2, Vicuna, Zephyr, Mistral) and show that X-ICL with ChatGPT few-shot yields over 6% improvement over ICL. Furthermore, while prompt selection strategies were previously shown to significantly improve ICL on in-distribution test sets, we show that these strategies do not match the efficacy of the X-ICL paradigm in robustness-oriented evaluations.