SFT: Truncate during dataset preparation, not collation#6155
Open
qgallouedec wants to merge 11 commits into
Open
SFT: Truncate during dataset preparation, not collation#6155qgallouedec wants to merge 11 commits into
qgallouedec wants to merge 11 commits into
Conversation
|
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. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0579163751
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Move sequence truncation in
SFTTrainerfrom the data collator into_prepare_dataset.Why
Truncation is a pure per-example slice. Doing it once during (cached) dataset preparation is cleaner than recomputing it in the collator on every batch, and keeps all dataset shaping (tokenize → build labels → truncate → pack) in one place.
It will also allow to drop rows with no trainable rewards, see #6025
Changes
_prepare_datasetnow truncatesinput_idsandlabelstomax_length(respectingtruncation_mode), right after labels are built. Skipped when packing (packing already chunks tomax_length).build_labelsnow drops thecompletion_mask/assistant_maskscolumns: they're fully baked intolabels.DataCollatorForLanguageModelingno longer truncates: themax_lengthandtruncation_modearguments are removed.With
skip_prepare_dataset=True, preparation (and therefore truncation) is skipped. The dataset must already be truncated.Note
Medium Risk
Changes default training data shape and breaks callers that relied on collator truncation or
skip_prepare_datasetwithout pre-truncated data; core SFT path behavior shifts for long sequences and masked labels.Overview
SFT sequence truncation moves from
DataCollatorForLanguageModelinginto_prepare_dataset, soinput_idsandlabelsare sliced tomax_length(withtruncation_mode) once during cached preparation—after label building, skipped when packing.DataCollatorForLanguageModelingno longer acceptsmax_lengthortruncation_mode; it only pads batches. Label building during preparation now dropsassistant_masks/completion_maskafter folding them intolabels.Behavior change: with
skip_prepare_dataset=True, truncation is not applied; datasets must already be length-limited. Assistant-only truncation regression (#3927) is asserted on the prepared dataset, not via the collator.Tests drop collator truncation cases and add preparation-focused coverage (
test_dataset_truncated_to_max_length, updated #3927 expectations).Reviewed by Cursor Bugbot for commit b3bbb4f. Bugbot is set up for automated code reviews on this repo. Configure here.