[WIP] MLM Training Objective #680
Draft
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
dlwh
reviewed
Aug 1, 2024
src/levanter/data/text.py
Outdated
| @@ -64,6 +65,65 @@ | |||
|
|
|||
| DEFAULT_IGNORE_INDEX = -100 # Mirrors pytorch's default ignore index | |||
|
|
|||
| class MaskedLmDataset(ShardableDataset[LmExample]): | |||
There was a problem hiding this comment.
fyi we're gonna do a big refactor on datasets soon, but I'll either handle the refactor or guide you through it)
src/levanter/data/text.py
Outdated
| def _create_mlm_example(tokens, key): | ||
| tokens_array = tokens.array | ||
|
|
||
| example = LmExample.causal(tokens=tokens, ignore_id=self.ignore_id) |
There was a problem hiding this comment.
you need a non-causal attention mask for Roberta, and you need to set a loss_mask to be only the masked tokens
There was a problem hiding this comment.
you also can't use the current LmExample actually because you need a separate targets field (with the non-masked tokens). With more work you could avoid the need for targets (with just masked tokens), but probably better to add an targets: Optional[NamedArray] to the class (or make your own class)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Introduces train_mlm.py, a new file adapted from train_lm.py, to support masked language modeling with dynamic masking as utilized in RoBERTa. A new class, MaskedLMDataset, has been implemented in text.py to handle dynamic masking. This class is instantiated and utilized within train_mlm.py, preserving all structural and sharding-related comments from the original train_lm.py to maintain clarity and continuity. The integration of MaskedLMDataset with the training script has been verified with appropriate parameters to ensure consistency with existing training workflows.