Closed xenova closed 1 month ago
Adds support for updated Llama 3 tokenizer.
Closes #739
Example JavaScript code:
import { AutoTokenizer } from "@xenova/transformers"; const tokenizer = await AutoTokenizer.from_pretrained("Xenova/llama3-tokenizer-new"); console.log(tokenizer.encode('hello world')); // [128000, 15339, 1917] console.log(tokenizer.encode('hello', 'world')); // [128000, 15339, 128000, 14957] console.log(tokenizer('hello world', { return_tensor: false })); // { input_ids: [ 128000, 15339, 1917 ], attention_mask: [ 1, 1, 1 ] } console.log(tokenizer('hello', { text_pair: 'world', return_tensor: false })); // { input_ids: [128000, 15339, 128000, 14957], attention_mask: [1, 1, 1, 1] } console.log(tokenizer('hello world', { return_token_type_ids: true, return_tensor: false })); // { input_ids: [128000, 15339, 1917], attention_mask: [1, 1, 1], token_type_ids: [0, 0, 0] } console.log(tokenizer('hello', { text_pair: 'world', return_token_type_ids: true, return_tensor: false })); // { input_ids: [128000, 15339, 128000, 14957], attention_mask: [1, 1, 1, 1], token_type_ids: [0, 0, 1, 1] }
Equivalent Python code:
from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained('Xenova/llama-3-tokenizer-new') print(tokenizer.encode('hello world')) # [128000, 15339, 1917] print(tokenizer.encode('hello', 'world')) # [128000, 15339, 128000, 14957] print(tokenizer('hello world')) # {'input_ids': [128000, 15339, 1917], 'attention_mask': [1, 1, 1]} print(tokenizer('hello', 'world')) # {'input_ids': [128000, 15339, 128000, 14957], 'attention_mask': [1, 1, 1, 1]} print(tokenizer('hello world', return_token_type_ids=True)) # {'input_ids': [128000, 15339, 1917], 'token_type_ids': [0, 0, 0], 'attention_mask': [1, 1, 1]} print(tokenizer('hello', 'world', return_token_type_ids=True)) # {'input_ids': [128000, 15339, 128000, 14957], 'token_type_ids': [0, 0, 1, 1], 'attention_mask': [1, 1, 1, 1]}
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
Adds support for updated Llama 3 tokenizer.
Closes #739
Example JavaScript code:
Equivalent Python code: