fix bias in eval per-char and per-byte normalization #6
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
kyleclo
approved these changes
Apr 2, 2025
kyleclo
left a comment
kyleclo
left a comment
There was a problem hiding this comment.
lgtm; can u update the changelog & make version update too in this PR
davidheineman
added a commit
to allenai/OLMo-core
that referenced
this pull request
Apr 4, 2025
Merge and update in-loop evals to match allenai/OLMo-in-loop-evals#6 Description from the `ai2-olmo-eval==0.7.1` PR: > our current in-loop does not match the OLMES standard (in our downstream evals), [the downstream normalization is here](https://github.com/allenai/oe-eval-internal/blob/main/oe_eval/metrics/metric.py#L229-L231). It appears the -1 term is an artifact from when we originally implemented in-loop evals. This will keep the original version of the metrics and add a version with the fixed normalization (`_v2`) --------- Co-authored-by: Dirk Groeneveld <[email protected]>
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Outdated suggestions cannot be applied.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.
Our in-loop evals subtract 1 from the denominator when normalizing by continuation length.
This means our current in-loop does not match the OLMES standard (in our downstream evals), the downstream normalization is here. It appears the
-1term is an artifact from when we originally implemented in-loop evals.This PR will match the normalization between in-loop to downstream.
Sanity Check
I have a gist copmaring the in-loop vs. downstream evals on
/oe-training-default/ai2-llm/checkpoints/mayeec/olmo-cookbook-1b-5xC-dclm-baseline-natural-9a234fde/step53971. For downstream taskarc_challenge:rc::olmes:full, this should match the in-loop keyarc_challenge_test_rc_5shot:https://gist.github.com/davidheineman/737ae4f50da7fe74c4e66d165025c68c: