These are the items in the vocab.json , sorted into separate categories: | |
1) Emojis - As you can see in the vocab.json , emojis are not represented in a comprehensible manner. | |
The text in the vocab json is not reversible, | |
If tokenization "😅" => ðŁĺħ | |
Then you can't prompt "ðŁĺħ" and expect to get the same result as the tokenized original emoji , "😅". | |
As such , a list of emojis must be fetched from elsewhere and placed alongside the vocab.json | |
2) Token type | |
A 'suffix' token is any token with a trailing whitespace </w> , these are the majority. | |
The tokens which lack a trailing whitespace </w> are sorted as 'prefix' tokens. | |
3) ID rarity | |
The higher the ID in the vocab.json , the less that word appeared in the CLIP training data. | |
The tokens are sorted into five groups : common (lowest IDs) => average => rare => weird => exotic (highest IDs) | |
4) Filtering (certain items are removed) | |
Tokens which contain special symbols that mess up the syntax have been removed. | |
For the full list of tokens refer to the vocab.json located in the 'raw' folder visible in the 'vocab folder'. | |