This paper introduces the ColorSwap dataset, designed to assess and improve the proficiency of multimodal models in matching objects with their colors. The dataset is comprised of 2,000 unique image-caption pairs, grouped into 1,000 examples. Each example includes a caption-image pair, along with a “color-swapped” pair. We follow the Winoground schema: the two captions in an example have the same words, but the color words have been rearranged to modify different objects. The dataset was created through a novel blend of automated caption and image generation with humans in the loop. We evaluate image-text matching (ITM) and visual language models (VLMs) and find that even the latest ones are still not robust at this task. GPT-4V and LLaVA score 72% and 42% on our main VLM metric, although they may improve with more advanced prompting techniques. On the main ITM metric, contrastive models such as CLIP and SigLIP perform close to chance (at 12% and 30%, respectively), although the non-contrastive BLIP ITM model is stronger (87%). We also find that finetuning on fewer than 2,000 examples yields significant performance gains on this out-of-distribution word-order understanding task.

Statements involving metalinguistic self-reference (“This paper has six sections.”) are prevalent in many domains. Can large language models (LLMs) handle such language? In this paper, we present “I am a Strange Dataset”, a new dataset for addressing this question. There are two subtasks: generation and verification. In generation, models continue statements like “The penultimate word in this sentence is” (where a correct continuation is “is”). In verification, models judge the truth of statements like “The penultimate word in this sentence is sentence.” (false). We also provide minimally different metalinguistic non-self-reference examples to complement the main dataset by probing for whether models can handle metalinguistic language at all. The dataset is hand-crafted by experts and validated by non-expert annotators. We test a variety of open-source LLMs (7B to 70B parameters) as well as closed-source LLMs through APIs. All models perform close to chance across both subtasks and even on the non-self-referential metalinguistic control data, though we find some steady improvement with model scale. GPT 4 is the only model to consistently do significantly better than chance, and it is still only in the 60% range, while our untrained human annotators score well in the 89-93% range. The dataset and evaluation toolkit are available at https://github.com/TristanThrush/i-am-a-strange-dataset

Evaluation is a key part of machine learning (ML), yet there is a lack of support and tooling to enable its informed and systematic practice. We introduce Evaluate and Evaluation on the Hub—a set of tools to facilitate the evaluation of models and datasets in ML. Evaluate is a library to support best practices for measurements, metrics, and comparisons of data and models. Its goal is to support reproducibility of evaluation, centralize and document the evaluation process, and broaden evaluation to cover more facets of model performance. It includes over 50 efficient canonical implementations for a variety of domains and scenarios, interactive documentation, and the ability to easily share implementations and outcomes. The library is available at https://github.com/huggingface/evaluate. In addition, we introduce Evaluation on the Hub, a platform that enables the large-scale evaluation of over 75,000 models and 11,000 datasets on the Hugging Face Hub, for free, at the click of a button. Evaluation on the Hub is available at https://huggingface.co/autoevaluate.

We introduce Dynatask: an open source system for setting up custom NLP tasks that aims to greatly lower the technical knowledge and effort required for hosting and evaluating state-of-the-art NLP models, as well as for conducting model in the loop data collection with crowdworkers. Dynatask is integrated with Dynabench, a research platform for rethinking benchmarking in AI that facilitates human and model in the loop data collection and evaluation. To create a task, users only need to write a short task configuration file from which the relevant web interfaces and model hosting infrastructure are automatically generated. The system is available at https://dynabench.org/ and the full library can be found at https://github.com/facebookresearch/dynabench.

Detecting online hate is a complex task, and low-performing models have harmful consequences when used for sensitive applications such as content moderation. Emoji-based hate is an emerging challenge for automated detection. We present HatemojiCheck, a test suite of 3,930 short-form statements that allows us to evaluate performance on hateful language expressed with emoji. Using the test suite, we expose weaknesses in existing hate detection models. To address these weaknesses, we create the HatemojiBuild dataset using a human-and-model-in-the-loop approach. Models built with these 5,912 adversarial examples perform substantially better at detecting emoji-based hate, while retaining strong performance on text-only hate. Both HatemojiCheck and HatemojiBuild are made publicly available.

In Dynamic Adversarial Data Collection (DADC), human annotators are tasked with finding examples that models struggle to predict correctly. Models trained on DADC-collected training data have been shown to be more robust in adversarial and out-of-domain settings, and are considerably harder for humans to fool. However, DADC is more time-consuming than traditional data collection and thus more costly per annotated example. In this work, we examine whether we can maintain the advantages of DADC, without incurring the additional cost. To that end, we introduce Generative Annotation Assistants (GAAs), generator-in-the-loop models that provide real-time suggestions that annotators can either approve, modify, or reject entirely. We collect training datasets in twenty experimental settings and perform a detailed analysis of this approach for the task of extractive question answering (QA) for both standard and adversarial data collection. We demonstrate that GAAs provide significant efficiency benefits with over a 30% annotation speed-up, while leading to over a 5x improvement in model fooling rates. In addition, we find that using GAA-assisted training data leads to higher downstream model performance on a variety of question answering tasks over adversarial data collection.

We introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. Dynabench runs in a web browser and supports human-and-model-in-the-loop dataset creation: annotators seek to create examples that a target model will misclassify, but that another person will not. In this paper, we argue that Dynabench addresses a critical need in our community: contemporary models quickly achieve outstanding performance on benchmark tasks but nonetheless fail on simple challenge examples and falter in real-world scenarios. With Dynabench, dataset creation, model development, and model assessment can directly inform each other, leading to more robust and informative benchmarks. We report on four initial NLP tasks, illustrating these concepts and highlighting the promise of the platform, and address potential objections to dynamic benchmarking as a new standard for the field.

We present the results of the first task on Large-Scale Multilingual Machine Translation. The task consists on the many-to-many evaluation of a single model across a variety of source and target languages. This year, the task consisted on three different settings: (i) SMALL-TASK1 (Central/South-Eastern European Languages), (ii) the SMALL-TASK2 (South-East Asian Languages), and (iii) FULL-TASK (all 101 x 100 language pairs). All the tasks used the FLORES-101 dataset as the evaluation benchmark. To ensure the longevity of the dataset, the test sets were not publicly released and the models were evaluated in a controlled environment on Dynabench. There were a total of 10 participating teams for the tasks, with a total of 151 intermediate model submissions and 13 final models. This year’s result show a significant improvement over the known base-lines with +17.8 BLEU for SMALL-TASK2, +10.6 for FULL-TASK and +3.6 for SMALL-TASK1.

Despite recent progress, state-of-the-art question answering models remain vulnerable to a variety of adversarial attacks. While dynamic adversarial data collection, in which a human annotator tries to write examples that fool a model-in-the-loop, can improve model robustness, this process is expensive which limits the scale of the collected data. In this work, we are the first to use synthetic adversarial data generation to make question answering models more robust to human adversaries. We develop a data generation pipeline that selects source passages, identifies candidate answers, generates questions, then finally filters or re-labels them to improve quality. Using this approach, we amplify a smaller human-written adversarial dataset to a much larger set of synthetic question-answer pairs. By incorporating our synthetic data, we improve the state-of-the-art on the AdversarialQA dataset by 3.7F1 and improve model generalisation on nine of the twelve MRQA datasets. We further conduct a novel human-in-the-loop evaluation and show that our models are considerably more robust to new human-written adversarial examples: crowdworkers can fool our model only 8.8% of the time on average, compared to 17.6% for a model trained without synthetic data.

We present a human-and-model-in-the-loop process for dynamically generating datasets and training better performing and more robust hate detection models. We provide a new dataset of 40,000 entries, generated and labelled by trained annotators over four rounds of dynamic data creation. It includes 15,000 challenging perturbations and each hateful entry has fine-grained labels for the type and target of hate. Hateful entries make up 54% of the dataset, which is substantially higher than comparable datasets. We show that model performance is substantially improved using this approach. Models trained on later rounds of data collection perform better on test sets and are harder for annotators to trick. They also have better performance on HateCheck, a suite of functional tests for online hate detection. We provide the code, dataset and annotation guidelines for other researchers to use.

Previous studies investigating the syntactic abilities of deep learning models have not targeted the relationship between the strength of the grammatical generalization and the amount of evidence to which the model is exposed during training. We address this issue by deploying a novel word-learning paradigm to test BERT’s few-shot learning capabilities for two aspects of English verbs: alternations and classes of selectional preferences. For the former, we fine-tune BERT on a single frame in a verbal-alternation pair and ask whether the model expects the novel verb to occur in its sister frame. For the latter, we fine-tune BERT on an incomplete selectional network of verbal objects and ask whether it expects unattested but plausible verb/object pairs. We find that BERT makes robust grammatical generalizations after just one or two instances of a novel word in fine-tuning. For the verbal alternation tests, we find that the model displays behavior that is consistent with a transitivity bias: verbs seen few times are expected to take direct objects, but verbs seen with direct objects are not expected to occur intransitively.