Model details can be found at https://huggingface.co/gpt2 Popular phrases such as "pair of" are examples of repetitive

In this work, we aim to advance our understanding by presenting a straightforward and fundamental explanation from the data perspective. Our preliminary investigation reveals a strong correlation between the degeneration issue and the presence of repetitions in training data. Subsequent experiments also demonstrate that by selectively dropping out the attention to repetitive words in training data, degeneration can be significantly minimized.

In Appendix A, we start by showing more qualitative results of chats and their retrieval results, and BLIP2 chats compared to a human answerer. Next, in Appendix B, we present the few shot instructional prompts that were used by different LLMs for generating follow-up questions. Another example in Figure 2 describes two trains, searched by the text "A train that is parked next to another train". Figure 3 demonstrates a case where the description "a small and dirty kitchen with pots and food everywhere" is ambiguous, subjective to the viewer and may match many images in the corpus. In Figure 4 we show an example of a dialog between ChatIR and a human.

Active Learning (AL) deals with identifying the most informative samples forlabeling to reduce data annotation costs for supervised learning tasks. ALresearch suffers from the fact that lifts from literature generalize poorly andthat only a small number of repetitions of experiments are conducted. To overcomethese obstacles, we propose CDALBench, the first active learning benchmarkwhich includes tasks in computer vision, natural language processing and tabularlearning. Furthermore, by providing an efficient, greedy oracle, CDALBenchcan be evaluated with 50 runs for each experiment. We show, that both thecross-domain character and a large amount of repetitions are crucial forsophisticated evaluation of AL research. Concretely, we show that thesuperiority of specific methods varies over the different domains, making itimportant to evaluate Active Learning with a cross-domain benchmark.Additionally, we show that having a large amount of runs is crucial. With onlyconducting three runs as often done in the literature, the superiority ofspecific methods can strongly vary with the specific runs. This effect is so strong, that, depending on the seed, even a well-established method's performance can be significantly better and significantlyworse than random for the same dataset.

Imbalanced classification, where one class is observed far less frequently than the other, often causes standard training procedures to prioritize the majority class and perform poorly on rare but important cases. A classic and widely used remedy is to augment the minority class with synthetic examples, but two basic questions remain under-resolved: when does synthetic augmentation actually help, and how many synthetic samples should be generated? We develop a unified statistical framework for synthetic augmentation in imbalanced learning, studying models trained on imbalanced data augmented with synthetic minority samples and evaluated under the balanced population risk. Our theory shows that synthetic data is not always beneficial. In a ``local symmetry" regime, imbalance is not the dominant source of error near the balanced optimum, so adding synthetic samples cannot improve learning rates and can even degrade performance by amplifying generator mismatch. When augmentation can help (a ``local asymmetry" regime), the optimal synthetic size depends on generator accuracy and on whether the generator's residual mismatch is directionally aligned with the intrinsic majority-minority shift. This structure can make the best synthetic size deviate from naive full balancing, sometimes by a small refinement and sometimes substantially when generator bias is systematic. Practically, we recommend Validation-Tuned Synthetic Size (VTSS): select the synthetic size by minimizing balanced validation loss over a range centered near the fully balanced baseline, while allowing meaningful departures when the data indicate them. Simulations and a real sepsis prediction study support the theory and illustrate when synthetic augmentation helps, when it cannot, and how to tune its quantity effectively.

There are a number of diverging hypotheses about the neural text degeneration problem, i.e., generating repetitive and dull loops, which makes this problem both interesting and confusing. In this work, we aim to advance our understanding by presenting a straightforward and fundamental explanation from the data perspective. Our preliminary investigation reveals a strong correlation between the degeneration issue and the presence of repetitions in training data. Subsequent experiments also demonstrate that by selectively dropping out the attention to repetitive words in training data, degeneration can be significantly minimized. Furthermore, our empirical analysis illustrates that prior works addressing the degeneration issue from various standpoints, such as the high-inflow words, the likelihood objective, and the self-reinforcement phenomenon, can be interpreted by one simple explanation. That is, penalizing the repetitions in training data is a common and fundamental factor for their effectiveness. Moreover, our experiments reveal that penalizing the repetitions in training data remains critical even when considering larger model sizes and instruction tuning.

Large-scale language models often learn behaviors that are misaligned with user expectations. Generated text may contain offensive or toxic language, contain significant repetition, or be of a different sentiment than desired by the user. We consider the task of unlearning these misalignments by fine-tuning the language model on signals of what not to do. We introduce Quantized Reward Konditioning (Quark), an algorithm for optimizing a reward function that quantifies an (un)wanted property, while not straying too far from the original model. Quark alternates between (i) collecting samples with the current language model, (ii) sorting them into quantiles based on reward, with each quantile identified by a reward token prepended to the language model's input, and (iii) using a standard language modeling loss on samples from each quantile conditioned on its reward token, while remaining nearby the original language model via a KL-divergence penalty. By conditioning on a high-reward token at generation time, the model generates text that exhibits less of the unwanted property. For unlearning toxicity, negative sentiment, and repetition, our experiments show that Quark outperforms both strong baselines and state-of-the-art reinforcement learning methods like PPO, while relying only on standard language modeling primitives.