Is the BOS token id of 128000 **hardcoded** into the llama 3.2 tokenizer?

[rasyosef]

System Info

• transformers version: 4.45.1
• Platform: Linux-5.15.154+-x86_64-with-glibc2.31
• Python version: 3.10.13
• Huggingface_hub version: 0.23.2
• Safetensors version: 0.4.3
• Accelerate version: 0.30.1
• Accelerate config: not found
• PyTorch version (GPU?): 2.1.2+cpu (False)
• Tensorflow version (GPU?): 2.15.0 (False)
• Flax version (CPU?/GPU?/TPU?): 0.8.4 (cpu)
• Jax version: 0.4.28
• JaxLib version: 0.4.28
• Using distributed or parallel set-up in script?:

Who can help?

@ArthurZucker @itazap

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

I trained the llama 3.2 tokenizer using an Amharic language corpus and a vocab size of 28k, but when I use it to tokenize text, the first token id is still 128000 when it should have been the new tokenizer's BOS token id of 0.

And here's a tokenization of an example text. As can be seen, the first token id is 128000 when it should have been 0.

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("rasyosef/llama-3.2-amharic-tokenizer-28k")

text = "ሁሉም ነገር"
inputs = tokenizer(text, return_tensors="pt")
print(inputs["input_ids"])

Output:

tensor([[128000,   1704,    802]])

Expected behavior

The first token id of the tokenized text should be the new tokenizer's BOS token id of 0 instead of the original llama 3.2 tokenizer's BOS token id of 128000. The vocab size is 28000 and the number 128000 should not appear anywhere in the input_ids list.

This is causing index out of range errors when indexing the embedding matrix of a newly initialized model.

[ArthurZucker]

Hey! Your post_processor is the following:

print(tokenizer._tokenizer.post_processor)
Sequence(processors=[ByteLevel(add_prefix_space=True, trim_offsets=False, use_regex=True), TemplateProcessing(single=[SpecialToken(id="<|begin_of_text|>", type_id=0), Sequence(id=A, type_id=0)], pair=[SpecialToken(id="<|begin_of_text|>", type_id=0), Sequence(id=A, type_id=0), SpecialToken(id="<|begin_of_text|>", type_id=1), Sequence(id=B, type_id=1)], special_tokens={"<|begin_of_text|>":SpecialToken(id="<|begin_of_text|>", ids=[128000], tokens=["<|begin_of_text|>"])})])

This is because the TemplateProcessor was not changed.
This is probably an issue with train_new_from_iterator is that what you used?

Otherwise, we can have an extra layer of logic in transformers that modifies the bos_token when loading if there is a post processor.

TLDR; your fix:

tokenizer._tokenizer.post_processor = Sequence( [ 
ByteLevel(add_prefix_space=True, trim_offsets=False, use_regex=True), 
TemplateProcessing(
    single="<|begin_of_text|> $0",
    pair="<|begin_of_text|> $A <|begin_of_text|> $B:1",
    special_tokens=[("<|begin_of_text|>", 0)],
)
])

this outpuded:

tensor([[   0, 1704,  802]])

[ArthurZucker]

(FYI @itazap )

[rasyosef]

Thanks @ArthurZucker , this fixed the issue. And yes, I used train_new_from_iterator.

[ArthurZucker]

Okay, in that case I think we ought to update the post_processor when we train from new!

[ArthurZucker]

Actually you can pass special tokens here:

transformers/src/transformers/tokenization_utils_fast.py

Lines 739 to 741 in c185608

	special_tokens_map (`Dict[str, str]`, *optional*):
	If you want to rename some of the special tokens this tokenizer uses, pass along a mapping old special
	token name to new special token name in this argument.

which already updates them. I am working on improving the api to make it simpler!

[rasyosef]

Does this affect the postprocessor? After training the tokenizer, the special token ids are correctly updated to numbers below the vocab size (0 for bos and 1 for eos). As you pointed out earlier, it was the bos token id in the postprocessor that was not updated and remained at 128000. Here are the first few lines of tokenizer_config.json

{
  "added_tokens_decoder": {
    "0": {
      "content": "<|begin_of_text|>",
      "lstrip": false,
      "normalized": false,
      "rstrip": false,
      "single_word": false,
      "special": true
    },
    "1": {
      "content": "<|end_of_text|>",
      "lstrip": false,
      "normalized": false,
      "rstrip": false,
      "single_word": false,
      "special": true
    },
    "2": {
      "content": "<|reserved_special_token_0|>",
      "lstrip": false,
      "normalized": false,
      "rstrip": false,
      "single_word": false,
      "special": true
    },

[rasyosef]

Hi @ArthurZucker , is there a way to add new tokens to a BPE tokenizer like Llama 3's while keeping the original vocabulary?

I wanted to extend the vocabulary of Llama 3.2 1B's tokenizer and then continue pretraining it on a new (Amharic language) corpus. tokenizer.add_tokens(['ĠáĭĭáĬĵ', 'ĠáĬ¨áī°áĪĽ']) does not seem to work, possibly because the list of merges need to be update as well.

UPDATE: I got it to work. I had to edit the tokenizer.json file of the Llama 3.2 tokenizer and add the new (Amharic) tokens and list of merges, and IT WORKS!

[ArthurZucker]

Hey! you can quite simply add the tokens, AS LONG AS you use the byte level normalizer and not the ``pre_tokenizer`!
It was added in: huggingface/tokenizers#1555

[rasyosef]

Cool, Thanks!

[csHuangfdu]

If I want to output tensor([[1704, 802]])
that is, I don't want to output bos_token
what should I do?

[ArthurZucker]

You can set add_bos_token=False and from_slow=False