training

Training Vocabularies

There are 4 steps to generating a vocabulary.

1. Prepare your dataset

To train a vocabulary, you need a dataset as a single plain text file. This dataset should represent exactly what you want to vocabulary to represent, and in the same proportions. For example, if you want the vocabulary to cover both English and French then you should ensure the dataset is 50% English and 50% French. Around 1GB is a reasonable dataset size for a large model, or 100-200MB for a small model.

2. Generate tokens

The next stage is to use the tool getalltokens to filter and export a list of all tokens that appear in the dataset. This is the stage at which you choose your charset and optimization mode settings. It takes around 30 minutes.

3. Train vocabulary

Using trainvocab, the tokens list is distilled down to your target vocabulary size. This takes 12 - 24 hours.

4. Export vocabulary

exportvocab then converts the output of trainvocab into a finished TokenMonster vocabulary file that can be used with the Python, Go & Javascript tokenizers. You can add special tokens and resize the vocabulary.

Let's go through each stage in more detail.

First compile or download the binaries

You need the getalltokens, trainvocab and exportvocab binaries. You can either download these precompiled for your architecture from here. Or you can build them yourself from source using Go 1.20 as follows:

Linux:

Install Golang:

apt install golang

Clone TokenMonster repository:

git clone github.com/alasdairforsythe/tokenmonster
cd tokenmonster/training

Build the binaries:

go mod init tokenmonster
go mod tidy
go build getalltokens.go
go build trainvocab.go
go build exporttokens.go

Windows

Download the Windows binaries directly from here. To use them, launch Windows Powershell and navigate to the directory where the executables are. Then type ./getalltokens. It should display the usage instructions, if not you probably need to add execute permissions (Google and ChatGPT both know how to do that.)

Prepare your dataset

The vocabulary is generated from the words, phrases and grammar in the dataset. It's important to prepare a dataset that properly represents what you want to tokenize. If a word is not in your dataset, it won't be in your vocabulary. Additionally, the importance of a word within the vocabulary is directly related to the frequency that it appears within the dataset. It's important therefore to ensure that different styles of writing, different languages, etc. that you intend to be included, make up a reasonable portion of the dataset. It is not necessary to distribute styles of text throughout the dataset; it's okay, for example, to have the first half of the file English, and the second half French.

I provide the datasets I used to generate the pretrained vocabularies on Hugging Face. You're welcome to use any of these to complement your dataset. Two helper Python scripts are included download_code_samples.py and extract_text_from_jsonl_parquet.py which I used to generate those datasets.

The diversity and range of the dataset depends upon whether it's intended use is specialized or general. Let's say, for example, that it's for a text-generation model that writes short stories in the style of Dr. Seuss when prompted with a subject. If that were the case, your dataset for the vocabulary can and should be the same dataset as used for training the model. In this case, you don't need to be worried about overfitting on the vocabulary because the output is always in the same style as the training data. On the other hand, if you are generating a vocabulary for a ChatGPT-like general model, you would want to be much more careful to avoid overfitting and likely use a different dataset to that used for training the model.

Generate tokens

Once you have your dataset ready as a single .txt file you can generate the tokens using getalltokens. This process takes from a minute to an hour, depending on the dataset size and how many threads you give it. For the pre-built vocabularies, -mode clean took around 20 minutes, and -mode strict took 2 minutes.

It's important to choose the correct settings, so I will go through them one by one.

Usage of ./getalltokens:
  -capcode int
        0 = disabled, 1 = deleteToken only, 2 = enabled (default 2)
  -charset string
        one of: UTF-8, none (default UTF-8)
  -chunk-size string
        the number of bytes processed at a time, higher is faster but requires more RAM (default 100MB)
  -dataset string
        filename of the dataset plain-text (required)
  -max-token-length int
        the maximum length of a token (default 40)
  -micro-chunks int
        the higher this number, the slower it is but it will reduce peak memory usage (default 5)
  -min-occur int
        tokens will be trimmed if they occur less frequently than this in the dataset (default 1 per 10MB)
  -min-occur-byte int
        single bytes will be trimmed if they occur less frequently than this in the dataset (default min-occur)
  -min-occur-chunk int
        tokens will be trimmed if they occur less frequently than this per chunk (default 4)
  -min-occur-micro-chunk int
        tokens will be trimmed if they occur less frequently than this per micro-chunk (default 2)
  -mode string
        0 = unfiltered, 1 = clean, 2 = balanced, 3 = consistent, 4 = strict (required)
  -norm string
        combine any of the following: NFD, lowercase, accents, quotemarks, collapse, trim, leadingspace, newlines (default NFD)
  -only-latin
        if enabled, tokens that contains letters must be in Latin script (default false)
  -only-valid
        if enabled, tokens must contain full and valid UTF-8 characters, except single byte tokens (default false)
  -output string
        output filename for the dictionary (required)
  -workers int
        number of worker threads to run (default 8)

-capcode

Capcode which is an alternative encoding for upper-casing, which eliminates the need for separate lowercase and uppercase variants of the same word. -capcode 2 is strongly recommended. For a language that doesn't use capitals, or for code only model, you can use level -capcode 1, which disables capcode but still uses the forward delete token. For languages or formats that don't use spaces as word separators, you can disable capcode completely with -capcode 0

-charset

Enter utf-8 if your dataset is in UTF-8 (probable) otherwise None. It affects some of the normalizations.

-norm

Options are: nfd lowercase accents quotemarks collapse trim leadingspace unixlines

For efficient, lossless tokenization of UTF-8 text, choose -norm NFD. All other normalizations are lossy. The parameters you choose here will be applied automatically to the vocabulary and to any string that is tokenized with the vocabulary.

• nfd: Applies NFD normalization to the UTF-8 text, which makes it more efficient to tokenize.
• lowercase: Converts to lowercase.
• accents: Removes accents, e.g. á → a
• quotemarks: Converts curly quotes to ASCII quotes ‘’“” → ''""
• collapse: Converts 2 or more sequential spaces to a single space (affects space character only)
• trim: Removes whitespace from the beginning and end.
• leadingspace: Adds a single space character at the beginning, if there isn't one already.
• unixlines: Converts /r/n to /n

These can be combined together, e.g. -norm "lowercase collapse trim quotemarks unixlines"

-only-latin

If enabled, tokens may not contain characters from non-Latin scripts. If you only intend to tokenize Latin script, it's best to enable this. Characters in non-Latin scripts can still be tokenized with single byte tokens.

-only-valid

If enabled, tokens may not contain invalid UTF-8. I recommend this in most cases.

-mode

The optimization mode is one of the most important parameters, as this completely changes the way your vocabulary works.

-mode unfiltered as described does not filter any tokens. Use this when your dataset is already clean, such as a text file that contains only gene sequences. Do not use this for natural language, and especially do not use it for code because it will result in overfitting.

-mode clean provides minimal normalization to avoid overfitting. Specifically, words are forced to begin with a space intead of end with one, and it restricts the amount of whitespace that can appear in a token alongside letters or numbers.

-mode balanced optimizes tokens for whole words. A token is restricted from covering one and a half words, or ending on a capcode modifier, with exceptions.

-mode consistent restricts the amalgamation of letters, numbers, delimiters, and punctuation within a single token. Some combinations are permitted, such as a word token ending with a comma or a space, or containing numbers. Most other combinations are now segregated into distinct tokens to maintain consistency. In balanced mode, a single word might have multiple representations in different tokens, e.g. then - then, then... However, in -mode=consistent there is a limited set of basic variations. Constraints also apply to open-closer delimeters, such as ([{'", and two open-closers within the same token is disallowed.

-mode strict attempts to have only 1 token for each word, however it is written. HELLO, "Hello" & hello! will all be tokenized with the same hello token, combined with capcode and punctuation tokens. It is allowed for tokens to cover multiple words, so how and how are you may be separate tokens. Open-closers such as ([{'" are restricted from being combined with other marks, with some exceptions.

As a rule of thumb, small models should use strict or consistent, medium models should use consistent or balanced, and large models should use balanced or clean. You can view the difference between them on the online viewer.

-max-token-length

The default and maximum is 40, I suggest leaving it at that, unless you have a good reason to limit it. Reducing max-token-length is unlikely to make the training or tokenization faster because the algorithm is optimized for -max-token-length 40.

-min-occur

Tokens that occur less frequently than this throughout the dataset will be deleted at the end. If you don't specify, it will be 10 for every 100MB of dataset size (e.g. 100 for 1GB).

-chunk-size, -micro-chunks, -min-occur-chunk, -min-occur-micro-chunk

When running getalltokens there is a tradeoff between speed, RAM and the pruning rate. To manage this, we have these parameters.

-chunk-size is the number of bytes that are processed before the tokens are pruned for minimum frequency of min-occur-chunk. Then within each chunk, the tokens are sorted and pruned -micro-chunks number of times, each time with tokens occurring less often than -min-occur-micro-chunk being deleted.

The default settings of -chunk-size 100MB -micro-chunks 5 is unlikely to use more than 8 - 16 GB of RAM, depending on the optimization mode (unfiltered & clean use more RAM because there are more tokens). If you find the RAM swapping a little it's okay, but if it's too much then kill the process and run it again with -micro-chunks 10.

If -micro-chunks 10 is still using too much RAM, you can use -chunk-size 10MB -min-occur-chunk 2 -micro-chunks 10 -min-occur-micro-chunk 1 which should use very little RAM but take a lot longer.

-min-occur-byte

This is how many times an individual byte must occur before it's allocated a token. By default this is the same as min-occur, but if you know your dataset is clean you might want to set it to 1 so that all individual bytes that occur will have a token that covers them. Or you might want to set it to 3 so that individual bytes that can occur have tokens, but a little bit of corrupt data in the dataset won't result in tokens being allocated.

Single-byte tokens are different to regular tokens in that they're not filtered out of the vocabulary. Single-byte tokens you count here will most likely end up in your final vocabulary, unless you can specifically exclude them at the next stage.

-workers

This is the number of threads used. More is faster, give it as many as you have available.

Train vocabulary

trainvocab trains the vocabulary on the tokens produced by getalltokens. Unlike getalltokens this process uses very little RAM, but it does take a long time. On 8 threads, it'll take between 12-24 hours to generate a final vocabulary. There is a -fast option that will produce a slightly less optimal vocabulary in about an hour, which is intended for testing the viability of a vocabulary before doing the full training.

Usage of ./trainvocab:
  -dataset string
        filename of the dataset plain-text (required)
  -dictionary string
        filename of the dictionary generated by getalltokens or any of the saved output files from this app (required)
  -dictionary2 string
        a second dictionary that will be merged with the first (optional)
  -dir string
        directory to save the results within (required)
  -exclude-other-bytes
        any single bytes not specifically included will not receive tokens, even if they were in the training dataset (default false)
  -fast
        runs 10x faster but the vocabulary might not be as optimal (default false)
  -include-128-bytes
        include tokens representing every ASCII character inc. control characters (default false)
  -include-256-bytes
        include tokens representing every possible byte (default false)
  -include-ascii-bytes
        include tokens for every printable ASCII character, inc. \r\n\t (default false)
  -include-extended-bytes
        include tokens for ASCII & UTF-8 chars used in English, e.g. “£©áê (default false)
  -include-missing-bytes
        add tokens for any single bytes found in the dataset that are not tokens already (default false)
  -include-utf8-bytes
        include tokens for every byte that can occur in UTF-8 text (default false)
  -keep-trying int
        program will exit when unable to find a better match this many times in a row (default 1000)
  -midway-target int
        beneath this the full dataset is used for every worker (default 6x vocab-size)
  -percentage int
        percentage of the dataset given to each worker before midway-target (default 15)
  -special string
        filename of a JSON file containing special tokens (optional)
  -vocab-size int
        vocabulary size, e.g. 32000 (required)
  -workers int
        number of worker threads to run, excluding main thread (default 8)

-dataset is the dataset. -dictionary is the tokens file from getalltokens. -vocab-size is your target vocabulary size. -workers is the number of threads to run (best to set it to 1 less than the number of CPU threads.)

-dir is the output directory. It's created if it does not exist. Many intermediate files are created so make sure it's an empty directory. Also trainvocab will attempt to resume/continue from it's previous state if the directory is not empty and it sees the intermediary files. That can be useful in case you need to stop and start it, but it means bad things will happen if you use the same -dir for different vocabularies.

Before running trainvocab you first need to decide on the parameters for what to do with single-byte tokens. Those are tokens for individual bytes, such standard English characters that are represented with ASCII, and starting or continuation bytes of multi-byte sequences in UTF-8.

Unlike regular tokens, all single-byte tokens are included in every vocabulary. In total there are 256 possible single-byte tokens. So for large vocabularies it's common to include all of them. This ensures that any data can be tokenized with the vocabulary, even binary data. However, that's a bit wasteful for smaller vocabularies, especially for specialized models or models that only generate ASCII text.

-include-256-bytes includes 256 tokens for all possible single bytes.

-include-128-bytes includes the full set of 128 ASCII character codes, including control characters.

-include-ascii-bytes includes all printable ASCII characters, as well as \r\n\t (newlines and tabs), but none of the control characters. This is what you want for small models that are only trained on English text.

-include-extended-bytes is the same as -include-ascii-bytes but it also adds the following: £€©®™°%¢¥—–•‘’“”áéíóúýàèìòùâêîôûäëïöüñãõçåæœ

-include-utf8-bytes gives a token for every byte that could make up any part of a valid UTF-8 character sequence. This is what you want if your model only generates UTF-8 text.

-include-missing-bytes dynamically adds tokens for all bytes discovered during training that don't already have tokens.

-exclude-other-bytes instructs trainvocab to ignore single-byte token from the tokens file generated by getalltokens. This is useful to include along with any of the above if your dataset is not clean, to ensure tokens are not allocated to characters that are not supposed to be there (for example, some UTF-8 characters within a dataset that's meant to be only ASCII.)

For the pretrained vocabularies, I used -include-256-bytes for 50256, 65536 & 100256 vocab sizes. I used -include-utf8-bytes -exclude-other-bytes for most of the others, and for 8000 and less vocabulary size I used -include-extended-bytes -exclude-other-bytes.

When trainvocab has finished, there will be hundreds of files in the -dir output directory. The one you want is the alphabetically first file, as this is the list of vocab-size tokens that tokenized the dataset with the least number of tokens. Once you've exported this as a vocabulary, you can delete this directory. The other files that begin with numbers are all the other vocabularies that were within 1% of the best vocabulary, and the other files are intermediary files that can be used to resume the training from any of those intermediary points.

Note that the file format of the tokens files in the output directory is the same format as the tokens generated by getalltokens. As such, you can use any of the files in the output directory as the input -dictionary for trainvocab. For example, you could use the final tokens from a 50,000 size vocabulary as the input dictionary for training a 10,000 size vocabulary. Doing so is not recommended because the optimal 10,000 tokens is not necessarily a subset of the optimal 50,000 tokens, but if you're in a hurry, you can do it.

-special

This parameter is used to add "special tokens". A special token is special because (1) it's always included in the vocabulary, and (2) regular tokens cannot contain it. Special tokens can be added during training, or at a later time.

To include special tokens during training, -special refers to a JSON file in the following format:

{ "special": [ "TOKEN1", "TOKEN2", "TOKEN3" ] }

Export vocabulary

Once a vocabulary has been generated it's not yet in the vocabulary format, it's still just a list of tokens. To convert it to a vocabulary for use with the tokenizing libraries, you use exportvocab.

The standard usage is straightforward:

./exportvocab -input directory -output file.vocab

This will automatically select the best vocabulary from directory (as generated by trainvocab) and output the final vocabulary to file.vocab.

Your vocabulary is now complete. However, you can do more with exportvocab. It can import and export to and from 3 formats: token files, TokenMonster vocabs, and YAML. For example, you can export a vocabulary in YAML format, order it with the highest scoring tokens at the top, and then open it with a text editor to see what the vocabulary looks like:

./exportvocab -input-vocab file.vocab -output-yaml file.yaml -order-by-score

You can also add tokens, resize the vocabulary, search the vocabulary, etc.

Usage of ./exportvocab:
  -add-single-bytes string
        enter "256", "128", "ascii", "extended" or "utf8" to add tokens for those individual bytes (optional)
  -add-special-token string
        a single special token to add to the vocabulary (optional)
  -delete-single-bytes
        deletes all the single byte tokens except those specified from add-single-bytes (optional)
  -exists string
        check if a token exists in the vocabulary (optional)
  -input string
        tokens file or directory from trainvocab, if directory it will load the best performing tokens file in the directory (optional)
  -input-vocab string
        an existing TokenMonster vocabulary file (optional)
  -input-yaml string
        filename of a YAML file containing modifications or a new vocabulary (optional)
  -order-by-score
        orders output-txt by token score (descending) instead of alphabetically (optional) (default false)
  -output string
        filename of the vocabulary to output (optional)
  -output-tokens string
        converts a vocabulary back to a tokens file that can be used with trainvocab (optional)
  -output-yaml string
        filename to export the vocabulary in YAML format (optional)
  -reset-token-ids
        resets the IDs of the tokens to be sequential from zero (optional) (default false)
  -resize int
        resizes the vocabulary to this many tokens by deleting the worst scoring tokens (optional)
  -unk string
        set to true or false to enable or disable the UNK token (optional)

-add-single-bytes & -add-special-token allow you to add single-byte or special tokens. If you combine this with -resize you can also keep the vocabulary within a target size. For example, to add a special token to an existing vocabulary of size 10000, but not increase the vocabulary size, you could use:

./exportvocab -input-vocab myvocab.vocab -add-special-token "<eos>" -resize 10000 -output mynewvocab.vocab -reset-token-ids

By default, token IDs are fixed, which means that if you resize or delete a token there will be gap in the token IDs. If you don't want this pass -reset-token-ids, which will assign new IDs to all the tokens alphabatically, beginning from zero.

-unk can be used to enable or disable the UNK token. If enabled, during tokenization, any byte for which there is no token will be covered with the UNK token. If disabled, a byte without a token is skipped. Vocabularies that used -include-256-bytes cannot have an UNK token because all bytes already have tokens.

-exists can be used if you want to check whether a token exists in a vocabulary. You should note that words generally begin with a space. Example:

./exportvocab -input-vocab myvocab.vocab -exists " cheesecake"

.