[LLaMA3] 'add_bos_token=True, add_eos_token=True' seems not taking effect

[kiva12138]

System Info

Platform = Windows
PyTorch = 2.3.0
Transformers = 4.41.0

Who can help?

No response

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

import torch
from transformers import AutoTokenizer

LLaMAPath = '/path/to/llama3-8b'

# The following two yields the same results, all of them contains BOS token and no EOS token
tokenizer = AutoTokenizer.from_pretrained(LLaMAPath, add_bos_token=True, add_eos_token=True)
# tokenizer = AutoTokenizer.from_pretrained(LLaMAPath, add_bos_token=False, add_eos_token=False)

tokenizer.add_special_tokens({"pad_token": "<|reserved_special_token_0|>"}) 
inputs = tokenizer(['hi, how are you today?'], padding=True, return_tensors='pt')
print(inputs)

All of the statements above produce [128000, 6151, 11, 1268, 527, 499, 3432, 30]

Expected behavior

I think when using tokenizer = AutoTokenizer.from_pretrained(LLaMAPath, add_bos_token=True, add_eos_token=True), we get [128000, 6151, 11, 1268, 527, 499, 3432, 30, 128001],

when using tokenizer = AutoTokenizer.from_pretrained(LLaMAPath, add_bos_token=False, add_eos_token=False), we get [6151, 11, 1268, 527, 499, 3432, 30],

[eyloncaplan]

I'm having the same issue. Neither of these change the encodings:
tokenizer.add_bos_token = False
tokenizer.add_eos_token = True

[amyeroberts]

cc @ArthurZucker

[ArthurZucker]

Hey! This is related to #30607, the tokenizer for Llama3 is a PreTrainedTokenizerFast, not the LLamaTokenizer or a LlamaTokenizerFast. Though it might actually be good to support an easy way to add bos and eos. Currently what you have to do is update the TemplateProcessor which is fairly annoying (not beginner friendly).

That's something which should be handle on the tokenizers side

[eyloncaplan]

Hey! This is related to #30607, the tokenizer for Llama3 is a PreTrainedTokenizerFast, not the LLamaTokenizer or a LlamaTokenizerFast. Though it might actually be good to support an easy way to add bos and eos. Currently what you have to do is update the TemplateProcessor which is fairly annoying (not beginner friendly).

That's something which should be handle on the tokenizers side

@ArthurZucker I think it's called TemplateProcessing, not TemplateProcessor. For those wondering this is how I used it to get the tokenizer to put the eos token:

bos = "<|begin_of_text|>"
eos = "<|end_of_text|>"
tokenizer._tokenizer.post_processor = processors.Sequence(
    [
        processors.ByteLevel(trim_offsets=False),
        processors.TemplateProcessing(
            single=f"{bos}:0 $A:0 {eos}:0",
            pair=f"{bos}:0 $A:0 {bos}:1 $B:1 {eos}:1",
            special_tokens=[
                (bos, tokenizer.bos_token_id),
                (eos, tokenizer.eos_token_id),
            ],
        ),
    ]
)

Now I'm worried that the padding tokens won't get added properly, but that's a different issue...

[ArthurZucker]

Padding token is unrelated, it's added if you ask the tokenizer to pad the input!
And yes, thanks for providing the snippet @eyloncaplan 😉