What are 'Tokens' ?

What are 'Tokens' in LLM's? What is the size of a Token?

In the context of Large Language Models (LLMs), tokens are the fundamental units of text that the model processes and understands. They can be thought of as building blocks that the LLM uses to construct and comprehend language.

Here's a breakdown of tokens in LLMs:


Tokens represent smaller pieces of text compared to whole words. This allows the LLM to handle complex language structures and account for variations in word forms (e.g., singular vs. plural, different tenses).

Types of Tokens:

Words: Often, individual words are used as tokens.

Subwords: In some cases, words are broken down into smaller units like prefixes, suffixes, or character n-grams (sequences of characters). This can be beneficial for handling rare words or recognizing morphological variations.

Special Tokens: Special tokens can be used to represent things like the beginning or end of a sentence, punctuation marks, or unknown words.

Token Size:

Variable: The size of a token can vary depending on the specific LLM and the tokenization technique used.

Common Sizes: Common token sizes include characters, words, or subwords of a fixed length (e.g., 3 characters).

Impact on Performance: The size and type of tokens can influence the LLM's performance in areas like vocabulary coverage and efficiency. Here's an analogy:

Imagine a child learning to build with Legos. Words would be like complete Lego bricks, while subwords would be smaller pieces like individual Lego studs. Tokens provide the LLM with a flexible set of building blocks to construct and understand language.

Additionally: The specific details of tokenization (how text is broken down into tokens) are often part of the LLM's architecture and might not be publicly available. The number of tokens in a piece of text is a crucial factor when considering the cost of training or using an LLM, as it influences the amount of data the model needs to process. I hope this explanation clarifies the concept of tokens in LLMs!


Here's an example of how a long sentence can be broken down into tokens using different tokenization techniques:

Sentence: "The quick brown fox jumps over the lazy dog." (10 words)

  1. Word-Level Tokens:

This is the simplest approach, where each word becomes a separate token.

Tokens: ["The", "quick", "brown", "fox", "jumps", "over", "the", "lazy", "dog", "."] (10 tokens)

  1. Subword-Level Tokens (Character Trigrams):

This method breaks down words into smaller pieces of 3 characters (trigrams). Special tokens (e.g., , ) might be added for sentence boundaries.

Tokens: ["", "The", "qui", "ick", "bro", "own", "fox", "jum", "mps", "ove", "r th", "e la", "zy d", "og", ""] (16 tokens)

  1. Subword-Level Tokens (BPE):

Byte Pair Encoding (BPE) is a more advanced technique that analyzes the training data to identify frequently occurring character pairs. These pairs are then merged into new tokens, creating a flexible vocabulary that can handle both common and rare words.

Tokens: (This would depend on the specific BPE model trained on a large corpus. Here's a hypothetical example): ["", "Th", "e", " ", "qu", "ick", " ", "br", "ow", "n", " ", "fo", "x", " ", "ju", "mp", "s", " ", "ov", "er", " ", "th", "e", " ", "la", "zy", " ", "do", "g", ""] (Possible tokens, might be more or less depending on the BPE model).

This example illustrates how tokenization can vary depending on the chosen method. The size and type of tokens can influence the LLM's efficiency and its ability to handle complex language structures.

Last updated