dbrx-base-tokenizer

Why should you use this and not the tiktoken included in the orignal model?

1. This tokenizer is validated with the https://huggingface.co/datasets/xn (all languages) to be encode/decode compatible with dbrx-base tiktoken
2. Original tokenizer pad the vocabulary to correct size with <extra_N> tokens but encoder never uses them
3. Original tokenizer use eos as pad token which may confuse trainers to mask out the eos token so model never output eos.
4. [NOT FIXED: INVESTIGATING] config.json embedding size of "vocab_size": 100352 does not match 100277

modified from original code @ https://huggingface.co/Xenova/dbrx-instruct-tokenizer

Changes:
1. Remove non-base model tokens
2. Keep/Add `<|pad|>` special token to make sure padding can be differentiated from eos/bos.
3. Expose 15 unused/reserved `<|extra_N|>` for use

# pad token
 "100256": {
      "content": "<|pad|>",
      "lstrip": false,
      "normalized": false,
      "rstrip": false,
      "single_word": false,
      "special": true
    },

# 15 unused/reserved extra tokens
"<|extra_0|>": 100261
"<|extra_1|>": 100262
...
"<|extra_14|>": 100275

DBRX Instruct Tokenizer

A 🤗-compatible version of the DBRX Instruct (adapted from databricks/dbrx-instruct). This means it can be used with Hugging Face libraries including Transformers, Tokenizers, and Transformers.js.

Example usage:

Transformers/Tokenizers

from transformers import GPT2TokenizerFast

tokenizer = GPT2TokenizerFast.from_pretrained('Xenova/dbrx-instruct-tokenizer')
assert tokenizer.encode('hello world') == [15339, 1917]

Transformers.js

import { AutoTokenizer } from '@xenova/transformers';

const tokenizer = await AutoTokenizer.from_pretrained('Xenova/dbrx-instruct-tokenizer');
const tokens = tokenizer.encode('hello world'); // [15339, 1917]