README.md

metadata

license: mit

TokenMonster

The documentation and code is available on Github alasdairforsythe/tokenmonster.

The pretrained vocabularies are all available for download here.

July 11: TokenMonster v1.1.1 has been released. The "420" prebuilt vocabularies are being released as they are completed, at a rate of around 10 per day.

Choose a dataset from:

• code
• english
• englishcode
• fiction

Choose a vocab size from:

• 1024
• 2048
• 4096
• 8000
• 16000
• 24000
• 32000
• 40000
• 50256
• 65536
• 100256

Choose an optimization mode from:

• unfiltered
• clean
• balanced
• consistent
• strict

For a capcode disabled vocabulary add:

• nocapcode

And finally add the version number:

• v1

Examples:

• fiction-24000-consistent-v1
• code-4096-clean-nocapcode-v1

There are two additional vocabularies:

• gpt2
• llama