Getting the most out of your tokenizer for pre-training and domain adaptation
Last updated
Copyright Continuum Labs - 2023
Last updated
Tokenization is often an understudied and neglected component in the development of models.
In this February 2024 paper, the authors highlight that most published works use a single tokenizer for all experiments, often borrowed from another model, without performing rigorous analysis or ablations to optimise the tokenization process.
Furthermore, when fine-tuning a pre-trained LLM for a specific task or domain, the tokenizer is generally kept unchanged, leading to sub-optimal performance and efficiency.
The authors argue that the size of the tokenizer's vocabulary, the pre-tokenization regular expression, and the training data used for the tokenizer can significantly impact the model's generation speed, effective context size, memory usage, and downstream performance.
To address this issue, the authors train specialised Byte-Pair Encoding (BPE) code tokenizers and conduct extensive ablations to study the impact of tokenizer design on the performance of LLMs for code generation tasks such as HumanEval and MBPP.
They provide recommendations for selecting appropriate tokenizer hyper-parameters and suggest switching the tokenizer when fine-tuning a pre-trained LLM.
The experiments are performed on models trained from scratch and on pre-trained models, verifying the applicability of their findings to a wide range of use-cases.
The authors find that when fine-tuning on more than 50 billion tokens, it is possible to specialise the tokenizer of a pre-trained LLM to obtain significant gains in generation speed and effective context size.
Training Data: Using data sampled from the target domain (e.g., code) will increase compression for that domain.
Pre-tokenization Scheme: The regular expression used to split the text before applying BPE affects compression. Splitting on whitespaces prevents BPE from merging across words, leading to shorter tokens and worse compression.
Vocabulary Size: A larger vocabulary size leads to higher compression but increases computational and memory costs.
Normalized Sequence Length (NSL): Measures the average tokenized sequence length of a tokenizer compared to a baseline (Llama tokenizer). An NSL of 0.75 means the tokenizer uses 25% fewer tokens on average.
Bytes per Token: Calculated by dividing the number of UTF-8 bytes by the number of tokens, providing another measure of compression.
The authors use the BPE tokenization algorithm, implemented with the HuggingFace tokenizers library, which supports regular expression-based pre-tokenization and handles special formatting characters better.
Unsurprisingly, training the tokenizer on data from the target domain (code, English, multilingual) improves compression for that domain. Training on a mix of all three leads to the best overall compression.
Compression vs. Vocabulary Size: Larger vocabularies improve compression, but gains diminish exponentially as the vocabulary size increases.
Inference Optimal Vocabulary Size: The authors calculate the optimal vocabulary size for inference time by considering the trade-off between compression gains and additional computation costs.
Memory Optimal Vocabulary Size: The authors derive an equation to find the memory-optimal vocabulary size, considering the model size, sequence length, batch size, and the memory savings from reduced attention cache size due to compression.
Biomedical Domain
Develop a tokenizer tailored for biomedical literature, such as research papers, clinical notes, and scientific reports.
Train the tokenizer on a large corpus of biomedical texts, incorporating domain-specific vocabulary, abbreviations, and naming conventions.
Use pre-tokenization schemes that preserve important biomedical entities, such as gene names, protein structures, and chemical compounds.
Integrate domain-specific knowledge bases or ontologies to improve tokenization accuracy and semantic understanding.
Legal and Regulatory Domain
Create a tokenizer specifically designed for legal documents, contracts, regulations, and legislative texts.
Train the tokenizer on a diverse corpus of legal texts, including case laws, statutes, and regulatory guidelines.
Implement pre-tokenization rules that preserve legal terminology, citations, and references to specific clauses or sections.
Incorporate legal ontologies and dictionaries to accurately tokenize complex legal phrases and terms.
Financial and Accounting Domain
Develop a tokenizer tailored for financial reports, accounting statements, and market data.
Train the tokenizer on a corpus of financial documents, including annual reports, balance sheets, and market analyses.
Implement pre-tokenization rules to handle financial notation, abbreviations, and numerical representations.
Integrate domain-specific knowledge bases or lexicons to accurately tokenize financial terms and concepts.
Cybersecurity Domain
Create a tokenizer specifically designed for cybersecurity logs, incident reports, and technical documentation.
Train the tokenizer on a corpus of cybersecurity-related texts, including system logs, vulnerability reports, and security guidelines.
Implement pre-tokenization rules to preserve technical terminology, IP addresses, and other relevant cybersecurity entities.
Integrate domain-specific knowledge bases or ontologies to accurately tokenize cybersecurity-related terms and concepts.
Social Media and Conversational Data
Develop a tokenizer tailored for social media data, such as tweets, online forums, and chat conversations.
Train the tokenizer on a diverse corpus of social media data, including slang, abbreviations, and internet-specific language.
Implement pre-tokenization rules to handle emoticons, hashtags, and other social media-specific constructs.
Incorporate language models or lexicons specific to social media and conversational data to improve tokenization accuracy.
Based on the experiments described in this section, here are the key conclusions and best practices regarding tokenizer construction and usage:
Changing the tokenizer of a pre-trained LLM during fine-tuning can have a negligible impact on downstream performance, provided that the fine-tuning is done on a sufficiently large amount of data (50 billion tokens or more).
The authors demonstrate that models fine-tuned with alternative tokenizers like GPT-4 and Punct can achieve competitive or even better performance compared to models using the original Llama tokenizer.
The experiments suggest that the vocabulary size of the tokenizer (within the tested range of 32k to 256k) has a minimal impact on the downstream performance of the LLM.
The authors found no statistically significant correlation between vocabulary size and performance metrics like Pass@1 and Pass@100 on code generation tasks.
Using techniques like Fast Vocabulary Transfer (FVT) to initialize the new tokenizer's embeddings from the pre-trained model leads to noticeable performance improvements compared to not using FVT.
Extending an existing tokenizer (e.g., Llama) by adding domain-specific tokens provides only small gains compared to using a completely different tokenizer like GPT-4.
While highly compressed tokenizers like Identity can offer significant compression benefits, they may result in deteriorated downstream performance on code generation tasks.
Tokenizers like Punct and GPT-4, which strike a balance between compression and preserving syntactic and semantic information, can achieve both better performance and better compression compared to the Llama tokenizer.
The authors demonstrate that their findings regarding tokenizer switching and its negligible impact on performance hold true for larger LLMs like Llama 2 7B when fine-tuned on a sufficient amount of data.
In this study, the authors investigated the impact of tokenizer design choices on the performance, compression, and efficiency of large language models (LLMs), with a focus on code generation tasks.
Their findings highlight the importance of carefully considering tokenization strategies, as they can significantly influence model capabilities and resource utilization.
Through extensive experimentation, they demonstrated that changing the tokenizer of a pre-trained LLM during fine-tuning can have a negligible impact on downstream performance, provided that the fine-tuning is conducted on a sufficiently large dataset (50 billion tokens or more).
This insight opens up opportunities for optimizing tokenizers for specific domains or tasks without sacrificing model accuracy.
Furthermore, their results suggest that the vocabulary size of the tokenizer, within a reasonable range (32k to 256k), has a minimal effect on the LLM's downstream performance. This finding allows for flexibility in balancing compression and memory/compute trade-offs based on the specific requirements of the application.
Overall, this study underscores the importance of carefully considering tokenization strategies in the development and fine-tuning of LLMs, as well as the potential for optimizing tokenizers to enhance model efficiency and domain-specific performance.
01.AI Yi series models (2023): Discusses the Yi series of large language models available on the Hugging Face Model Repository, highlighting their advanced capabilities in language understanding.
Ahmad et al. (2021): Explores unified pre-training methods for program understanding and generation, demonstrating a method to enhance model efficiency in understanding and generating programmatic content.
Ainslie et al. (2023): Describes training generalized multi-query transformer models from multi-head checkpoints, advancing the flexibility of transformer architectures in handling diverse queries simultaneously.
Allal et al. (2023): Introduces 'Santacoder', a concept for a playful, yet robust approach to coding assistance, enhancing code-related tasks without aiming for overly ambitious, unattainable goals.
Almazrouei et al. (2023): Discusses the Falcon series of open language models, emphasizing the development of accessible and robust language model frameworks.
Anthropic (2023): Details the release of 'Claude' by Anthropic, focusing on a new language model that aims to improve ethical considerations and robustness in AI.
Austin et al. (2021): Investigates the use of large language models for program synthesis, showing their potential in automating coding tasks and generating programmatic content from high-level descriptions.
Biderman et al. (2023): Presents 'Pythia', a suite designed for analyzing large language models across different stages of their training and scaling, aiming to understand their behavior and improve their design.
Black et al. (2022): Discusses GPT-NeoX-20B, an open-source autoregressive language model, contributing to the open research and development of scalable language models.
Chen et al. (2021): Evaluates large language models trained on code, providing insights into their effectiveness and areas for improvement in programming language understanding.
Chirkova & Troshin (2023): Explores subtokenization options for pretraining large language models on source code, aiming to optimize model performance on coding tasks.
Deci (2023): Introduces 'Decicoder', a model touted as a new standard in efficient and accurate code generation, enhancing the capabilities of AI in software development.
DeepSeek AI (2023): Details a series of code language models, enhancing the tools available for developers and programmers in automated code generation.
Devlin et al. (2019): Describes the training of BERT, a foundational model that uses deep bidirectional transformers for improved language understanding, setting a new standard in NLP.
Elsen et al. (2023): Announces the release of Persimmon-8B, a language model designed to follow short-form instructions effectively, demonstrating its practical applications.
Forsythe (2023): Introduces 'Tokenmonster', a tokenizer and vocabulary trainer designed to improve the efficiency of language processing in Python, Go, and JavaScript.
Fried et al. (2023): Presents 'Incoder', a generative model for code infilling and synthesis, aimed at enhancing automated coding tasks by filling in gaps and generating syntactically correct code snippets.
Gee et al. (2022, 2023): Discusses methods for fast vocabulary transfer and multi-word tokenization for language model compression and sequence compression, respectively, aiming to enhance model efficiency and manage larger vocabularies effectively.
Gowda & May (2020): Investigates the optimal vocabulary size for neural machine translation, providing insights that help improve translation accuracy and efficiency.
Goyal et al. (2023): Describes the training of language models with pause tokens, introducing a method to enhance natural language generation by incorporating thoughtful pauses in speech or text.
guidance-ai (2023): Details a guidance language designed for controlling large language models, emphasizing the development of more responsive and controllable AI systems.
Jiang et al. (2023): Discusses 'Mistral 7b', a model designed for multi-turn program synthesis, enhancing the interactive capabilities of AI in coding tasks.
Kocetkov et al. (2022): Presents 'The Stack', a large dataset of permissively licensed source code, aimed at facilitating the training and development of code-oriented AI models.
Kudo (2018): Investigates subword regularization techniques, aiming to improve translation models by managing subword variability effectively.
Kudo & Richardson (2018): Introduces 'Sentencepiece', a tokenizer that simplifies text processing by providing a consistent subword tokenization method
