Add text2text evaluator#261
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
| if tokenizer is not None and feature_extractor is not None: | ||
| logger.warning("Ignoring the value of the preprocessor argument (`tokenizer` or `feature_extractor`).") | ||
| if pipe.task != self.task: | ||
| if (pipe.task != self.task) and not (self.task == "translation" and pipe.task.startswith("translation")): |
There was a problem hiding this comment.
Because for translation the task in the pipeline is translation_xx_to_yy where the placeholders are the languages.
|
That's great! We should include the new subclasses in https://github.com/huggingface/evaluate/blob/main/docs/source/package_reference/evaluator_classes.mdx I think, and probably mention them in the how-to guide. |
ola13
left a comment
ola13
left a comment
There was a problem hiding this comment.
Looks great, some comments and questions inline!
Co-authored-by: helen <31600291+mathemakitten@users.noreply.github.com>
…d-text2text-evaluator
|
Thanks for your feedback, I think I integrated everything. Let me know if you think anything else needs to be changed! |
mathemakitten
left a comment
mathemakitten
left a comment
There was a problem hiding this comment.
I think this looks good to me, just some minor changes needed to the examples in the doc; see comments.
@NimaBoscarino have you had a chance to try this out yet? We can merge this for now but make updates to text2text-generation if it doesn't work with HONEST.
Co-authored-by: helen <31600291+mathemakitten@users.noreply.github.com>
There was a problem hiding this comment.
Added a couple small comments! I don't think that this works with HONEST for the moment for a couple reasons, but I'm wrapping up a few tests and will post my thoughts + questions as soon as I'm done.
EDIT: After talking with Leandro, it sounds like HONEST would likely work better with a FillMaskEvaluator, so I can hack one together for that tomorrow!
Co-authored-by: Nima Boscarino <nima.boscarino@gmail.com>
* add text2text evaluator * add tests * Apply suggestions from code review Co-authored-by: helen <31600291+mathemakitten@users.noreply.github.com> * final tweaks to t2t evaluators * fix compute docstring in base class * add new split/subset logic * Apply suggestions from code review Co-authored-by: helen <31600291+mathemakitten@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: Nima Boscarino <nima.boscarino@gmail.com> * update docs Co-authored-by: helen <31600291+mathemakitten@users.noreply.github.com> Co-authored-by: Nima Boscarino <nima.boscarino@gmail.com>
6 participants
This PR adds three evaluators following the pipeline philosophy:
While they are functionality similar they use slightly different defaults. The parity tests for the Trainer are missing but we can add those in a separate PR I think.
(for some reason I can't add you as reviewer @fxmarty but I think you can comment anyway)
cc @philschmid