Closed NISH1001 closed 1 year ago
Re:
I found the issue. It's during the AccuracyForLanguageGeneration._tokenize(...)
process which is stripping off some texts. such as when both the predictions and references are just literal '$'
:
Re: I was able to patch it under try/catch block here: https://github.com/NISH1001/jury/commit/6bdf6800f7498b970fa8a81f3ea4bf88f6a12c32
Should I send a PR? I don't know if need to just throw a warning or also show the original <value>
for either of the pred/ref.
Hi @NISH1001, thanks for the heads-up, and also thanks for your comments, it is appreciated. I'll look into the PR asap.
Describe the bug I was running
RobertaForQuestionAnswering
on HuggingFace's squad-v2 train sets (~86k). TheAccuracy
metric atAccuracyForLanguageGeneration._compute_single_pred_single_ref
threw division by zero error.To Reproduce
datasets
squad-v2train
set.pipeline("question-answering", ...)
Expected behavior Run without error.
Exception Traceback (if available) If applicable, add full traceback to help explain your problem.
Environment Information:
evaluate==0.2.2
datasets==2.11.0
Thanks. Appreciate jury to exist. I could patch this by cloning and doing in-depth trace analysis. But, I wanted to know if there's a better way to patch this.