v2 code updates #236
Merged
v2 code updates #236
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
natolambert
approved these changes
May 31, 2025
rewardbench/generative_v2.py
Outdated
|
|
||
| ###Feedback: """ | ||
|
|
||
| AUTOJ_COARSE_SCORE_RUBRIC = """ |
There was a problem hiding this comment.
If not run for v2, I'd remove all the code we aren't using right now so we start with a cleaner file.
| self.tokenizer = tokenizer | ||
|
|
||
| def __call__(self, samples, **kwargs): | ||
| _ = kwargs.get("batch_size", 1) |
There was a problem hiding this comment.
If only tested with batch size 1, we need to raise a not implemented error if the user passes batch size >1
There was a problem hiding this comment.
Now tested with batch size 4, works fine!
Closed
This was
linked to
issues
Jun 2, 2025
Closed
Closed
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Outdated suggestions cannot be applied.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.
V2 Code Updates
Big updates are summarized as follows:
scripts/run_v2.pynow has support for processing the Ties subset, which uses a different scoring function over the whole Ties subsetrewardbench/utils.pyhas new functionsprocess_single_modeland helper functionsample_statsfor Ties scoring.reroll_and_score_datasetinutils.pyto now reasonably penalize for 2/3/4-way ties for top score— also generally changed this function from best-of-4 to more flexible num completions to accommodate Tiesrun_v2.pyalso has a bunch of miscellaneous small updates that I copied over from doing a diff withrun_rm.py— things like device map, handling bfloat16, quantization etc. They were missing becauserun_v2.pywas initially based on the olderrun_bon.py.rewardbench/generative_v2.pyseparate since each prompt and function is different, but could consolidate, just not sure how much benefit that gives. Happy to iterate. The generative v2 scripts now:utils.load_eval_dataset_multi, similar toutils.load_eval_datasetfrom v1 but more flexible for multiple chosen/rejected--score_w_ratingsargumentrun_generative_v2.pyenforces this.Testing
I have tested this branch's code with the image
saumyam/rewardbench-2-pr-0530-tie-penaltyon:Small things still to fix
generative_v2.py, but they are just as long ingenerative.py, which doesn't have issues?