Skip to content

[SFTTrainer] Fix the sequence length check of SFTTrainer#512

Merged
younesbelkada merged 4 commits intomainfrom
fix-sft-trainer-length-bug
Jul 12, 2023
Merged

[SFTTrainer] Fix the sequence length check of SFTTrainer#512
younesbelkada merged 4 commits intomainfrom
fix-sft-trainer-length-bug

Conversation

@younesbelkada
Copy link
Copy Markdown
Contributor

What does this PR do?

Fixes #467
Replaces #481

The original code lead to having examples that were filtered out by the previous check. This ultimately lead to having examples being ignored when processing the data. Probably a copy-pasta from: https://huggingface.co/learn/nlp-course/chapter7/6?fw=pt where I borrowed the original code

The fix is to remove completely that check

cc @lvwerra

@younesbelkada younesbelkada requested a review from lvwerra July 12, 2023 12:48
@HuggingFaceDocBuilderDev
Copy link
Copy Markdown

HuggingFaceDocBuilderDev commented Jul 12, 2023

The documentation is not available anymore as the PR was closed or merged.

@lvwerra
Copy link
Copy Markdown
Member

lvwerra commented Jul 12, 2023

We should also set padding=False and return_length=False. Otherwise looks good.

@younesbelkada younesbelkada merged commit f323090 into main Jul 12, 2023
@younesbelkada younesbelkada deleted the fix-sft-trainer-length-bug branch July 12, 2023 13:25
yxliu-TAMU pushed a commit to mincheolseong/ECEN743-GRPO-Project-Proposal that referenced this pull request Apr 20, 2025
…face#512)

* fix the sequence length check of `SFTTrainer`

* forward contrib credits from initial contribution

* forward contrib credits from initial contribution

* final comments

---------

Co-authored-by: mrm8488 <mrm8488@users.noreply.github.com>
Co-authored-by: BramVanroy <BramVanroy@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Trainer filters out any data != max length

5 participants