manu / tok-fr-en-code

Custom Tokenizer

Examples

Example sentence: This is a test sentence. On va voir comment elle est gérée .... 123 + 56 = 2567. Let's go! Imagine I have code 4 spaces. and a backslash!! Eléonore est un prénom français. __name__ isInstance

Encoded sentence: ['▁This', '▁is', '▁a', '▁test', '▁sent', 'ence.', '▁On', '▁va', '▁voir', '▁comment', '▁elle', '▁est', '▁g', 'érée', '▁....', '▁', '1', '2', '3', '▁+', '▁', '5', '6', '▁=', '▁', '2', '5', '6', '7', '.', "▁Let's", '▁go', '!', '▁Im', 'ag', 'ine', '▁I', '▁have', '▁code', '▁', '▁', '▁', '▁', '4', '▁spaces', '.\n', '▁and', '▁a', '▁', '▁', '▁', '▁', '▁', '▁back', 'sl', 'ash', '!!', '▁El', 'éon', 'ore', '▁est', '▁un', '▁prénom', '▁français.', '▁__name__', '▁is', 'Instance']

Decoded sentence: <s> This is a test sentence. On va voir comment elle est gérée .... 123 + 56 = 2567. Let's go! Imagine I have code 4 spaces. and a backslash!! Eléonore est un prénom français. __name__ isInstance

Usage

from transformers import LlamaTokenizerFast

tok = LlamaTokenizerFast.from_pretrained('<tok_name>')

tok.tokenize('This is a test sentence')

Dataset Stats

Samples are trained on dataset manu/tok-corpus-shuffled

The dataset consists of french, english and code samples

More info on the dataset can be found here

For speed purposes, the tokenizer was trained on a sample of the dataset. Only the first samples were selected.

Sample size: 5000000

Size of Sampled: 19.0 GB

Tokenizer Configs

Build from scratch: True

Pretrained tokenizer: None

Tokenizer is trained with digit separation, whitespaces (for code), byte fallback...