drop episodes with 0 advantages or truncated #580
Merged
Conversation
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
meta-cla
bot
added
the
CLA Signed
Summary:
Episodes with all rewards = 0 or =1 does not help learning as advantage would be 0. also, episodes with generations that are tuncated due to max_res_tokens would mostly get 0 rewards unnecessary as most of answers are at the end.
Dropping these episodes provides trainer better batches to learn from (at the cost of sampling efficiency)
{F1983571844}
{F1983571853}
Reviewed By: casteryh
Differential Revision: D87243621
Intaik
force-pushed
the
export-D87243621
branch
from
b9ca0cd to
95b82c8
Compare
casteryh
approved these changes
Nov 18, 2025
Contributor
casteryh
left a comment
casteryh
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #580 +/- ##
=======================================
Coverage 84.10% 84.10%
=======================================
Files 29 29
Lines 3687 3687
=======================================
Hits 3101 3101
Misses 586 586 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
daniellepintz
pushed a commit
to daniellepintz/torchforge
that referenced
this pull request
Nov 20, 2025
Summary: ## Bug Description: meta-pytorch#580 had incorrect indentation cuasing the input_ids, episodes varibles to be deleted inside the episodes building loop, causing program to hang. Next diff shall make background thread crashes to be surfaced to the main thread so that we know what thread crashed for what reason. Reviewed By: daniellepintz Differential Revision: D87554570
daniellepintz
pushed a commit
to daniellepintz/torchforge
that referenced
this pull request
Nov 20, 2025
Summary: ## Bug Description: meta-pytorch#580 had incorrect indentation cuasing the input_ids, episodes varibles to be deleted inside the episodes building loop, causing program to hang. Next diff shall make background thread crashes to be surfaced to the main thread so that we know what thread crashed for what reason. Reviewed By: daniellepintz Differential Revision: D87554570
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Summary:
Episodes with all rewards = 0 or =1 does not help learning as advantage would be 0. also, episodes with generations that are tuncated due to max_res_tokens would mostly get 0 rewards unnecessary as most of answers are at the end.
Dropping these episodes provides trainer better batches to learn from (at the cost of sampling efficiency)
{F1983571844}
{F1983571853}
Differential Revision: D87243621