Train sampler unification#43138
Conversation
|
This is my first PR, so I'm not too familiar with this process, but there's a natural consequence that this would outdate a lot of documentation. Should I go ahead and update all previous mentions of |
29e5f0b to
be3a2c2
Compare
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
@jiosephlee @SunMarc |
|
@romitjain I agree that this will make the Trainer more flexible. There are, however, general hesitations to adding arguments and bloating Trainer even further. Thoughts on adding an additional argument called |
|
View the CircleCI Test Summary for this PR: |
|
@SunMarc should I merge? |
|
@ArthurZucker I would appreciate your opinion on this |
|
Pls merge @ArthurZucker
Maybe in the future when users request this ! They can also overwrite the |
|
plz merge @ArthurZucker |
Resolve conflicts keeping train_sampling_strategy in place of group_by_length, adapting to upstream's reorganized docstrings. Co-Authored-By: Claude Opus 4.6 <[email protected]>
Remove 13 duplicate field definitions in TrainingArguments that existed in both our old layout and upstream's reorganized sections. Fix exaone_moe model card date per utils/add_dates.py. Co-Authored-By: Claude Opus 4.6 <[email protected]>
Head branch was pushed to by a user without write access
b5a3002 to
391b4d3
Compare
|
@ArthurZucker should be ready to merge again |
|
@SunMarc Did we not want a deprecation period/notice on removing this from TraningArgs? |
|
Do you think this is a widely used args ? Happy to add it back otherwise. This PR was supposed to be merged before v5 but it got forgotten. |
What does this PR do?
Adds ability to control the sampling strategy of data during training, unifying
group_by_lengthwith the default random sampling, and additionally adds sequential sampling. Following up on previous PR #42265@SunMarc