Skip to content

Random MLM replacement can sample special tokens #28

Description

@sbauer0

Random MLM replacement can sample special tokens

Hi, I noticed a small issue in the MLM masking logic.

In methylbert/src/data/vocab.py, the vocabulary is constructed with the special tokens prepended before the actual k-mer tokens:

special_tokens = ["<pad>", "<unk>", "<eos>", "<sos>", "<mask>"]

self.pad_index = 0
self.unk_index = 1
self.eos_index = 2
self.sos_index = 3
self.mask_index = 4

self.itos = list(special_tokens) + vocabs
self.stoi = {t: i for i, t in enumerate(self.itos

So the first normal k-mer token starts at index 5.

In methylbert/src/data/dataset.py, lines 192–195, the 10% random replacement step samples from the full vocabulary:

# 10% of the time, we replace masked input tokens with random word
indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
random_words = torch.randint(len(self.vocab), labels.shape, dtype=torch.int16)
inputs[indices_random] = random_words[indices_random]

Because torch.randint(len(self.vocab), ...) samples from all token IDs, it can also sample the special-token IDs 0–4, meaning a masked k-mer may be randomly replaced with <pad>, <unk>,<eos>, <sos>, or <mask>.

I think the random replacement step should probably sample only from the actual k-mer vocabulary and exclude special tokens. For example:

num_special_tokens = 5

random_words = torch.randint(
    low=num_special_tokens,
    high=len(self.vocab),
    size=labels.shape,
    dtype=torch.int16,
)
inputs[indices_random] = random_words[indices_random]

Happy to open a PR if this looks right.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions