Open BunnyNoBugs opened 1 year ago
Here the problem is related to the fact that the prompt format of the pretrained prompt "konodyuk/prompt_rugpt3large_detox_russe"
is defined as <P*X>{toxic_comment}<P*X>
. If it was <P*X>{text}<P*X>
, then the code would work, because {text}
is a default placeholder. For non-default interpolations you need to provide a dictionary containing the interpolation key. In this case the code should be
for i in tqdm(valid_dataset):
options = ppln(
i,
...
or
for i in tqdm(valid_dataset):
options = ppln(
{"toxic_comment": i["toxic_comment"]},
...
for more verbosity.
I think that I mistakenly uploaded a prompt checkpoint with the wrong format at some point, so it has nothing to do with datasets
version.
Will keep this issue open to fix the notebook when I get a chance. Thank you for the report.
Thank you for your answer, now the inference works properly!
Hello!
I was testing the library for detoxification, particularly your Inference from pretrained prompt (RUSSE Detox) tutorial. For me, inference returns a
KeyError
in Colab:Maybe the problem is with the specific version of
datasets
library needed? The notebook is written as if it was already installed, but I had to install it extra (the latest version).