Abstract:Cross-tokenizer distillation (CTD), the transfer of knowledge from a teacher to a student language model when the two use different tokenizers, remains a largely unsolved problem. Existing approaches rely on heuristic strategies to align mismatched vocabularies, introducing considerable complexity. In this paper, we propose a simple but effective baseline called Byte-Level Distillation (BLD) which enables CTD by operating at a common interface across tokenizers: the byte level. In more detail, we convert the teacher's output distribution to byte-level probabilities, attach a lightweight byte-level decoder head to the student, and distill through this shared byte-level interface. Despite its simplicity, BLD performs competitively with--and on several benchmarks surpasses--significantly more sophisticated CTD methods, across a range of distillation tasks with models from 1B to 8B parameters. Our results suggest that the byte level is a natural common ground for cross-tokenizer knowledge transfer, while also highlighting that consistent improvements across all tasks and benchmarks remain elusive, underscoring that CTD is still an open problem.
Abstract:It has been experimentally demonstrated that humans are able to learn in a manner that allows them to make predictions on categories for which they have not seen any examples (Malaviya et al., 2022). Sucholutsky and Schonlau (2020) have recently presented a machine learning approach that aims to do the same. They utilise synthetically generated data and demonstrate that it is possible to achieve sub-linear scaling and develop models that can learn to recognise N classes from M training samples where M is less than N - aka less-than-one shot learning. Their method was, however, defined for univariate or simple multivariate data (Sucholutsky et al., 2021). We extend it to work on large, high-dimensional and real-world datasets and empirically validate it in this new and challenging setting. We apply this method to learn previously unseen NLP tasks from very few examples (4, 8 or 16). We first generate compact, sophisticated less-than-one shot representations called soft-label prototypes which are fitted on training data, capturing the distribution of different classes across the input domain space. We then use a modified k-Nearest Neighbours classifier to demonstrate that soft-label prototypes can classify data competitively, even outperforming much more computationally complex few-shot learning methods.