Linting Xue, Aditya Barua, Noah Constant
Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel
2021 · arXiv
Most widely-used pre-trained language models operate on sequences of tokens corresponding to word or subword units. By comparison, token-free models that operate directly on raw text (bytes or characters) have many benefits: they can process text in any language out of the box, they are more robust to noise, and they minimize technical debt by removing complex and error-prone text preprocessing pipelines. Since byte or character sequences are longer than token sequences, past work on token-free models has often introduced new model architectures designed to amortize the cost of operating directly on raw text. In this paper, we show that a standard Transformer architecture can be used with minimal modifications to process byte sequences. We characterize the trade-offs in terms of parameter count, training FLOPs, and inference speed, and show that byte-level models are competitive with their token-level counterparts. We also demonstrate that byte-level models are significantly more robust to noise and perform better on tasks that are sensitive to spelling and pronunciation. As part of our contribution, we release a new set of pre-trained byte-level Transformer models based on the T5 architecture, as well as all code and data used in our experiments.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
6 selected excerpts · 6 total mentions
There are many approaches around this, the simplest being to treat bytes as tokens (cf: Google's ByT5[1]). Also, BLT[2] from Meta and ByteFormer[3] from Apple. [1]: https://arxiv.org/abs/2105.13626 [2]: https://arxiv.org/abs/2412.09871 [3]: https://arxiv.org/abs/2306.00238
> GPT-3 was terrible at jokes. Many LLMs are because it's hard to model puns in a stochastic parrot GPT-3 was great at jokes. The Navy Seals were hilarious (https://gwern.net/gpt-3#navy-seals). And the difficulty of modeling puns has nothing to do with 'stochastic parrots' and has everything to do with tokenization (https://gwern.net/gpt-3#bpes), in the same way that all those hyberbolic takes about how image-generation models were 'fundamentally flawed' because they couldn't do good text in images turned out to be BS and solely a matter of tokenization - drop in a character-tokenized LLM instead, even an obsolete & dumb one, and it instantly works (https://arxiv.org/abs/2105.13626#google).
This is a common myth, which I've written about before. https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que... The closest anyone's come to proving that byte-level tokenization is better is the ByT5 paper https://arxiv.org/abs/2105.13626 But they only showed evidence for improvement on specific tasks, not general performance, which is an important distinction. And their own benchmarks show that the improvements tend to be marginal: https://i.imgur.com/6Cw0APS.png
The ByT5 paper shows significantly better resilient to typos and names in common benchmarks, in addition to better reasoning capabilities with these types of tasks: https://arxiv.org/pdf/2105.13626.pdf The reason these models aren’t as popular is because inference speed is significantly slower for a given completion length, and training for the same context window in characters is significantly more expensive (I say this as someone who’s deployed ByT5 vs normal T5 in prod for this exact reason)
> LLMs work on tokens, which, for the most part, are words (longer words are generally split up into multiple tokens). Although this is mostly true, there are exceptions to the rule. Look at the ByT5 family[1] of models for instance. These models operate on byte level representations of text and are inherently token free. I read a recent paper[2] that found that replacing token based text encoders in text to image models with these byte level models improves the spelling abilities of the txt2img models (which are otherwise notoriously bad at spelling in all human languages). [1]: https://arxiv.org/abs/2105.13626 [2]: https://arxiv.org/abs/2212.10562
> The alternatives are learning at the character level (way more complex No, BPEs are more complex: you have a whole additional layer of preprocessing, with all sorts of strange and counterintuitive downstream effects and brand new ways to screw up (fun quiz question: everyone knows that BPEs use '<|endoftext|>' tokens to denote document breaks; what does the string '<|endoftext|>' encode to?). BPEs are reliably one of the ways that OA API users screw up, especially when trying to work with longer completions or context windows. But a character is a character. > and scales badly in memory/compute) Actually very competitive: https://arxiv.org/abs/2105.13626#google (Especially if you account for all the time and effort and subtle bugs caused by BPEs.)
More from the same topics.
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan +11
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Luca Beurer-Kellner, Beat Buesser, Ana-Maria Creţu +11
As AI agents powered by Large Language Models (LLMs) become increasingly versatile and capable of addressing a broad spectrum of tasks, ensuring their security has become a critical challenge.
Timo Schick, Jane Dwivedi-Yu, Roberto Dessì +5
Language models (LMs) exhibit remarkable abilities to solve new tasks from just a few examples or textual instructions, especially at scale.